# Market Research Market Research Report - Global

**Generated on:** 2026-10-13 13:44:37.752303  
**Industry:** Market Research  
**Geography:** Global  
**Details:** # Deep Research Prompt: Capability Discovery and Output Management in an Agentic AI Research Assistant ("Studio")

## How to use this prompt

Paste this in full into each model. For Grok, prefix it with: "Be maximally contrarian. Challenge every assumption below. What do users secretly hate about AI assistant side panels? What are the failure modes nobody talks about? What wrong assumptions do product teams make about one-click prompt buttons?"

---

## Context

**The product.** Compass is an agentic AI research assistant used by enterprise market research and insights teams. Users chat with it to create and launch qualitative research (video surveys, live interviews), analyse transcripts and recordings, and produce outputs. It can already generate documents, presentations, charts, infographics, interactive web reports, podcasts, video summaries and showreels, and can run multi-step guided flows (for example creating a research project step by step with human-in-the-loop checkpoints).

**The interface today.** A standard chat layout: left sidebar listing chat history, centre chat panel. Saved outputs (documents, uploads) currently live in the same sidebar as chat history under a heading like "My Saved Documents". There is an interim fix in place but it is acknowledged as temporary.

**The problem.**
1. **Capability discovery.** Users cannot see what Compass can do. Even internal staff have discovered major capabilities (such as building interactive web reports) by accident months after launch. If internal staff do not know, customers certainly do not.
2. **Output management.** Saved outputs are mixed in with chat history in the sidebar. Users find this confusing. Outputs are not grouped by the research project they came from, though the backend already supports linking assets to projects. Customers have said they want outputs grouped by project and may want their own folders.
3. **Saving ideas, not just quotes.** Session replays of users show they want to capture ideas, hypotheses and observations as they work, not only verbatim quotes. There is no home for this.

**The proposed solution: "Studio".** A panel in Compass, modelled on NotebookLM's Studio panel, that keeps chat in one place and everything else in another. It would contain:

- **Create** tab of one-click buttons, each starting its own guided flow: Create Research, Build a Presentation, Write a Report, Build an Interactive Report, Build an Infographic, Build a Chart, Generate a Podcast, Create a Video Summary, Create a Showreel, Write a Discussion Guide.
- **Analyse** tab of one-click buttons: Executive Summary, Pull Key Quotes, Generate Recommendations, Theme Breakdown, Sentiment Overview, Compare Segments, Build a Persona, Draft an Email Summary, Map the Journey, Find Archetypes, Trend Over Time.
- **A home for saved outputs** (documents, showreels, interactive reports, uploads), grouped by project.
- **Candidate additions:** an idea notebook, saved prompts and routines (scheduled or repeatable analyses), persona creation, account-level reports (for example a research gap analysis across all projects, or a trend radar), and topical digest buttons.

**Design work so far.** Prototypes exist for the Create Research flow (a step-by-step guided flow with confirmation checkpoints), the Studio panel as a third column, a sidebar rework separating outputs from chats, and routines. None are approved yet.

**Internal feedback and tensions to resolve.**
- A senior engineer is not convinced a third column is justified given typical end-user screen sizes, and would prefer the functionality folded into the existing left and centre panels. The designer has proposed that Studio could replace the chat panel on small screens, or sit on the left.
- Mobile behaviour is unresolved.
- What happens when a user clicks a button ten messages into an existing chat is unresolved. Does it start a new chat, inject into the current one, or open a sub-thread?
- For large project sets, the preference is a conversational human-in-the-loop step rather than filters.
- Naming is open. "Studio" itself is contested (alternatives floated: Lab, Workshop, Explorer). The saved-outputs area has had many names proposed (My Documents, Library, Vault, Creations, Outputs, My Work, or no blanket term at all, just the item types).
- Which buttons go in v1 is undecided, who writes and tests the button prompts is undecided, and the team agrees customers should be asked what they have used, would use and would not use before new buttons are built.

**Constraints.** Enterprise B2B. Users are insights professionals, often not technical. Many work on laptops with modest screens and sometimes in split-screen alongside other tools. The product must feel trustworthy and research-grade, not gimmicky.

---

## Research sections

Answer every section. Where you cite a product, cite the specific behaviour you observed or that is documented, not a general impression.

### 1. Capability discovery in AI assistants
- What patterns do leading AI assistants (ChatGPT, Claude, Gemini, Copilot, NotebookLM, Perplexity, Notion AI, Cursor) use to make capabilities discoverable: suggestion chips, slash commands, side panels, onboarding tours, "what can I do" prompts, templates galleries, contextual nudges?
- Which of these have published usage data or research showing they actually increase feature adoption, and which are decorative?
- What is known about the "empty box" problem (users not knowing what to type) and what has measurably reduced it?
- How do assistants surface capabilities contextually, for example offering "build a chart" only once data is present, versus always-visible menus? Which performs better and why?
- How do products handle the tension between discoverability and clutter when the capability list is long (20+ actions)?

### 2. One-click prompt buttons and guided flows
- What is the evidence on fixed prompt buttons versus free-text chat for task completion, error rate and user confidence?
- How do products decide which actions become buttons and which stay as free-text? What selection criteria and usage thresholds are used?
- What is the best-documented approach to writing, versioning and A/B testing the hidden prompts behind such buttons? Who should own them (design, product, research, prompt engineering)?
- When a button launches a multi-step guided flow with checkpoints, what interaction patterns keep the user oriented (progress indication, ability to go back, ability to break out into free chat)?
- What happens, in well-designed products, when a user triggers a templated action partway through an existing conversation? Document the options (new thread, inline, branched thread, side panel) and their trade-offs.

### 3. Panel architecture and layout
- Document how NotebookLM's three-panel layout (Sources, Chat, Studio) behaves across screen widths, including how it collapses, and what users and reviewers say about it.
- What other products use a three-column layout for AI work (for example Cursor, Claude Artifacts, ChatGPT Canvas, Gemini Canvas, Notion AI) and how do they handle the third panel at 1280px and 1366px widths, which are common enterprise laptop resolutions?
- What is the evidence for and against a persistent third panel versus a collapsible one versus folding actions into the sidebar and chat composer?
- What are the documented UX costs of panels that push the chat area below roughly 600px wide?
- How do these products behave on mobile and tablet? Which patterns degrade gracefully and which simply hide functionality?

### 4. Output and artefact management
- How do AI assistants with generative outputs (Claude Artifacts, ChatGPT Canvas, Gemini, Copilot, NotebookLM, Gamma, Perplexity Pages) store, list, name, version and let users return to outputs?
- Which products separate outputs from chat history, and which keep them together? What do users say about each?
- What are the best patterns for grouping outputs by the project or source they came from, while still allowing user-defined folders or tags? Where do these two models conflict?
- How do products handle outputs that are regenerated or edited: versions, overwrites, duplicates?
- How do products handle uploads sitting alongside generated outputs? Should they be the same list or distinct?

### 5. Idea capture and research notebooks
- What are the strongest patterns for capturing ideas, hypotheses and observations during analysis in research tools (Dovetail, EnjoyHQ, Condens, Notably, Marvin, Aurelius, NotebookLM notes, Obsidian, Readwise)?
- How do these tools link a saved idea back to its evidence (the quote, clip or transcript moment) without forcing structure on the user?
- What is the evidence on lightweight capture (one click, no form) versus structured capture (tags, categories, required fields) for adoption and later retrieval?
- How do tools make saved ideas useful later: resurfacing, clustering, feeding into reports? Which of these actually get used?
- Is an idea notebook better as part of the outputs area, part of the chat, or its own space?

### 6. Saved prompts, routines and repeatable analyses
- How do products let users save and re-run prompts (ChatGPT custom GPTs and scheduled tasks, Claude Projects, Gemini Gems, Perplexity Spaces, Zapier, n8n)?
- What is the uptake evidence for scheduled or recurring AI tasks in B2B tools? Who actually uses them?
- What interaction patterns make routines understandable to non-technical users (naming, previews, dry runs, notification of results)?
- Should saved prompts and routines live with the action buttons, with the outputs, or separately?

### 7. Naming and information architecture
- What do leading products call their action panels and their output areas? Compile a table of actual labels used in production.
- What does research on labelling say about generic container names (Library, Vault, Workspace) versus item-type labels (Reports, Presentations, Uploads) for findability?
- Which names carry unwanted connotations in enterprise B2B (for example "Lab" implying experimental, "Vault" implying locked away)?
- How should the information architecture signal the difference between chat history, generated outputs, user uploads and captured ideas without four separate navigation sections?

### 8. Enterprise insights and market research context
- What tools do insights and market research teams already use for producing and storing deliverables (Dovetail, Qualtrics, Remesh, UserTesting, Discuss.io, Canvs, Relative Insight, Notably), and how do their output and library areas work?
- What outputs do insights professionals actually produce most often and share upward? Which of the proposed Create and Analyse buttons map to real, frequent jobs and which look like capability demonstrations?
- What do insights teams expect around output governance: naming conventions, project association, sharing permissions, audit trail?
- What is known about how non-technical enterprise users respond to agentic, multi-step AI flows, particularly trust, checkpointing and the fear of the AI "doing too much"?

### 9. Mid-conversation actions and context handling
- When a user invokes a templated action inside a long chat, how do products decide what context the action receives (whole chat, last N messages, selected items, the project)?
- What patterns let users scope an action to a subset of data (specific projects, segments, date ranges) conversationally rather than through filter UI? Where has this been done well?
- How do products avoid the user losing their place in the chat when an action spawns a long output?

### 10. Measurement and success criteria
- What metrics do teams use to measure capability discovery (feature adoption curves, time to first use of each capability, proportion of users who use more than N capabilities)?
- What benchmarks exist for AI assistant feature adoption in B2B SaaS?
- How should the team instrument a Studio-style panel to learn which buttons earn their place and which should be removed?
- What qualitative signals from session replays and analytics would indicate the output management problem is solved?

### 11. Customer research design
- Design a short customer research study (interviews, card sort, tree test, or survey) that would tell the team which proposed buttons customers have used, would use, and would not use, and what they would call the outputs area. Specify method, sample, duration and the questions to ask.
- What research methods are best for validating navigation and naming decisions quickly with enterprise B2B users who have limited time?
- What are the common mistakes teams make when asking users which features they want, and how should the study avoid them?

### 12. Risks, failure modes and contrarian view
- What are the documented failure modes of NotebookLM-style Studio panels and prompt-button galleries? Where have products added them and later removed or redesigned them?
- What are the risks of shipping 21 buttons in v1? What is the evidence for a smaller set?
- What assumptions in the brief above are most likely to be wrong?
- What would a simpler solution look like that solves the three stated problems without a new panel at all?
- What does the team risk by waiting on customer research before building anything, and what does it risk by not waiting?

---

## Output instructions

- Provide specific, implementable recommendations, not ranges or "it depends". Where a decision is genuinely contingent, state the condition and the recommendation for each case.
- Cite sources for every factual claim about a product's behaviour, published research, or usage data. Include URLs. Flag where you are inferring rather than citing.
- For every recommendation, state the trade-off being accepted.
- Give a direct recommendation on each open question: panel placement (third column, left, folded in), mobile behaviour, mid-chat button behaviour, naming for the panel and the outputs area, which buttons belong in v1, and who should own the button prompts.
- Produce a proposed v1 scope: the smallest set of changes that addresses capability discovery, output confusion and idea capture, with a rationale for each inclusion and exclusion.
- End with a section titled "What surprised me" listing anything in your findings that contradicts the brief or common assumptions.
- Use UK English. Do not use em dashes.

---

# Compass Studio: Discover Capabilities Without Losing Research Control

## Executive summary

**Recommendation: do not ship a permanent Studio column.** Ship an **Actions** launcher beside the chat composer, a project-first **Library** in the existing left navigation, and an editable **Save idea** action. Open an output editor only when someone is working on an output. This addresses the three problems in the brief without making users learn a second application inside Compass.

- **Separate Conversations From Work**: ChatGPT now documents a separate Library, while Claude documents an Artifacts area accessible across conversations. Durable work does not have to remain buried in chat history. -> Give Compass outputs stable identities and project-based retrieval. **Trade-off:** another destination, but not another permanently visible column. [12][19]
- **Expose Capabilities Without Exposing Everything At Once**: NotebookLM combines source-grounded chat with one-click creation; ChatGPT also exposes Canvas through contextual activation, a composer tool and a slash command. -> Offer six prominent actions plus a searchable directory of every implemented capability. **Trade-off:** less immediate exposure for specialist actions, without making them undiscoverable. [13][10]
- **Do Not Confuse Availability With Adoption Evidence**: Published studies support helping people formulate and evaluate prompts, but do not establish that a Studio panel or a particular button count increases enterprise adoption. -> Measure completed research jobs, not button clicks. **Trade-off:** slower attribution, stronger evidence. [15][30]
- **Preserve The Evidence Chain**: Dovetail embeds highlights in documents to connect findings to raw data; ESOMAR asks suppliers about validation, data quality, lineage and human oversight. -> Every generated finding should retain accessible supporting evidence and its source scope. **Trade-off:** additional metadata and review work. [56][60]
- **Capture Thinking Before Classifying It**: Research on 27 knowledge workers identifies lightweight entry and unconstrained content as unmet needs; NotebookLM explicitly supports notes containing users' own thoughts. -> Save an idea immediately, with optional evidence and optional later classification. **Trade-off:** messier initial organisation. [34][7]
- **Treat Buttons As Product Code**: OpenAI's current prompting guidance recommends code-managed prompts, representative fixtures, evaluation checks and release controls. -> Product owns each action; research owns methodological acceptance; engineering owns versioning and deployment. **Trade-off:** buttons take longer to release than copy-pasted prompts. [4]
- **Do Not Copy Mobile Feature Gaps**: NotebookLM documents mobile limitations for notes, reports and data tables; ChatGPT Canvas documentation still lists mobile as forthcoming. -> Preserve Compass retrieval, idea capture and task review on mobile, even if complex editing becomes a full-screen view. **Trade-off:** fewer simultaneous panes. [8][10]
- **Separate Reuse From Autonomy**: Gems, project instructions and saved prompts reuse a method; scheduled tasks additionally introduce timing, notifications and approval. -> Defer scheduling until users repeatedly rerun a successful analysis. **Trade-off:** delayed automation, reduced operational risk. [38][29][25]
- **Research While Building Reversible Foundations**: The strongest initial case is for project association, output retrieval and idea capture, not a branded panel. -> Build those foundations while running the proposed two-week customer study. **Trade-off:** some implementation may change, but the team avoids both paralysis and a large speculative launch.

**Scope:** global enterprise market research and insights. Product behaviours below are documented behaviours, not claims of hands-on testing at specified screen widths. All Compass layouts, sample sizes, thresholds and launch gates are **proposals**, not industry benchmarks.

## 1. Capability discovery: combine a visible directory with contextual help

### What leading assistants actually expose

| Product | Documented discovery pattern | Implication for Compass |
|---|---|---|
| ChatGPT | Canvas can open when the task warrants it; users can request it, choose it from the composer toolbox or invoke `/canvas`. [10] | Provide both explicit access and relevant suggestions. |
| Claude | Users can request an artefact in a conversation or start from a template in the **Artifacts** tab. [19] | A capability can have a stable destination and a conversational entry point. |
| Gemini | Canvas is selected from the composer controls; an existing Canvas document offers **Create** transformations such as slides, an infographic or a web page. [11] | Offer transformations after a suitable source exists. |
| Microsoft Copilot | **Prompt Gallery** separates **Suggested**, **Your Prompts** and **Teams**. Saved and shared prompts are supported. [46] | Distinguish vendor-curated actions from personal and team methods. |
| NotebookLM | **Chat** offers suggested options; **Studio** lists source-based outputs. [49] | A source-aware action catalogue makes possibilities concrete. |
| Perplexity | The documented **Projects** area consolidates searches, tasks, files and instructions; an existing session can become a project. [9] | Discoverability can be organised around ongoing work rather than output formats. |
| Notion AI | Entry points include the left sidebar, a page-corner control and a desktop keyboard shortcut. [1] | Multiple routes can serve occasional and frequent users without multiple permanent panels. |
| Cursor | Reusable commands appear after typing `/`; Plan mode introduces an explicit reviewable planning workflow. [23][55] | Use commands for repeat users, but give non-technical users a labelled entry point. |

The common pattern is **multiple entry points into the same capability**, not one universally successful panel. Compass should retain free text while making its supported jobs visible.

### Case study: NotebookLM makes the input-output relationship visible

Google's redesign separates **Sources**, **Chat** and **Studio**. Sources hold project information, Chat discusses it with citations, and Studio transforms it into outputs. The documented design also allows the source viewer and notes editor to expand alongside one another. Its important contribution is the visible relationship between evidence and work, not simply the presence of a right-hand column. [13]

For Compass, copy that relationship: an action should show **what it will use** and **what it will produce**. Do not infer that NotebookLM's adoption proves a third column caused adoption. An XDA reviewer describes Studio as a "secret second app", which is useful qualitative commentary but hardly evidence that a separate panel solves discovery for everyone. [32]

### What the evidence does and does not establish

The "empty box" problem includes not knowing what the system can do, how to phrase a request, and how to judge success. In *Why Johnny Can't Prompt*, novices could iterate locally but struggled with robust evaluation; none used the supplied systematic testing interface. A separate task-support study found benefits from prompt suggestions, perception articulation and conversation explanations, but bundled those interventions rather than isolating a button gallery. [15][30]

**Decision:** show six stable action cards in a new-chat welcome state, an always-available **Actions** launcher, and no more than two contextual suggestions after relevant work. The two-suggestion limit is a design proposal. Show **Build a chart** after a usable table exists; keep the capability visible in the directory with "Requires structured data" when it does not. **Trade-off:** modest catalogue maintenance in return for avoiding both a blank box and a wall of buttons.

No reviewed source establishes a causal enterprise-adoption uplift for onboarding tours, "What can I do?" prompts or a persistent Studio gallery. Treat them as hypotheses, not as proven mechanisms or automatically "decorative" features.

## 2. One-click actions: standardise the workflow, not the user's judgement

### Evidence on buttons versus free text

| Evidence | What it supports | What it cannot establish |
|---|---|---|
| *Why Johnny Can't Prompt*: 10 participants with little or no prompt-design experience. [15] | Users need help with robustness and evaluation, not just wording. | A numerical task-completion advantage for fixed buttons. |
| Task-supportive LLM interaction: 16 participants completing 29 non-warm-up tasks. [30] | Context articulation and supportive interaction can help users manage complex tasks. | Which individual UI component caused the benefit. |
| PromptCharm: a small text-to-image study comparing an assisted interface with alternatives. [37] | Mixed-initiative editing can help in a bounded creative task. | Transferable error rates or confidence gains for market research. |

There is no defensible universal percentage improvement for "buttons versus chat" here. Compass should test **the same task, sources and output rubric** with free text and with the guided action.

### Case study: Cursor makes planning an inspectable object

Cursor's Plan mode explores the codebase, can ask clarifying questions, and produces a plan that opens as an editable virtual file. The user explicitly selects **Build** when satisfied. The documented workflow separates understanding the task from executing it. [55]

Compass should adopt that separation for **Create research**: **Objective -> Audience and method -> Discussion guide -> Review and launch**. Show the current step, completed decisions, Back, Save draft and "Ask Compass about this". Free-chat clarification should update the same draft rather than restart the wizard. Require a final explicit approval for launching research, contacting participants or incurring external cost. **Trade-off:** extra interaction before consequential actions, but not before every harmless draft edit.

### Which jobs deserve buttons?

Use a button when the job has a recognisable name, recurring demand, a clear input contract, an evaluable result and a safe recovery path. Leave exploratory, unusual and highly bespoke questions in chat. Competitors' internal selection thresholds are not publicly established by the reviewed material; the thresholds in section 10 are proposed Compass rules.

A button should initiate a **named workflow**, not silently paste an opaque instruction. Before execution, show the selected project, source count, audience, output format and any cost or publication consequence. Permit editing these values.

### Ownership, versioning and testing

OpenAI's current guidance says to treat prompts as application code, review changes with product behaviour, run representative tests and use release controls. It explicitly recommends code-managed production prompts rather than creating new reusable prompt objects. [4]

**Recommendation:** make the product manager accountable for each action. The research lead signs off analytical validity; the designer owns wording and state transitions; the AI engineer owns implementation, evaluation and rollback. Record action version, model version, retrieval settings, output schema and test-set version together. Test missing evidence, multilingual data, contradictory participants, sparse segments and unsupported requests before release. **Trade-off:** shared ownership creates coordination overhead, but avoids making a designer's hidden prompt an ungoverned research method.

For live A/B tests, vary one component at a time, use the same eligible population and score outputs blind to variant. Do not optimise solely for faster generation or user confidence: confidently wrong findings are worse than cautious drafts.

## 3. Panel architecture: preserve workspace before adding a third column

### What is documented, and what is not

| Product | Documented desktop arrangement | Mobile/tablet qualification | Exact behaviour at 1280/1366px |
|---|---|---|---|
| NotebookLM | Sources, Chat and Studio; adaptable source-and-note arrangement. [13] | Mobile Studio supports several formats, but notes, reports and data tables are documented gaps. [8] | No verified breakpoint specification in the reviewed documentation. |
| Claude | An artefact opens beside the conversation. [19] | Its help describes creating, viewing and editing artefacts on iOS and Android. [19] | Not specified. |
| ChatGPT Canvas | Requested content opens in a right-hand window. [10] | Documentation lists web, Windows and macOS; mobile is described as forthcoming. [10] | Not specified. |
| Gemini Canvas | A Canvas panel supports editing and transformations. [11] | Shared Canvas links cannot be opened in the Gemini mobile app. [11] | Not specified. |
| Notion AI | Sidebar entry can open a full-page view; a page-corner entry also exists. [1] | The reviewed source does not establish complete tablet/mobile parity. | Not specified. |
| Cursor | An Agent panel and an editable planning file support the workflow. [55] | A desktop coding workflow is not evidence of a suitable mobile research layout. | Not specified. |

These products are not all equivalent "three-column layouts". Some reveal a work surface only when required. Do not turn unspecified competitor breakpoints into invented specifications.

### Case study: NotebookLM's mobile omissions expose the cost of imitation

NotebookLM's mobile app preserves source selection and Studio generation for several formats, but its help explicitly says notes, mind maps, reports and data tables are not yet available. This is a concrete example of a multi-surface desktop product reducing functionality on mobile rather than simply rearranging it. [8]

Compass's idea-capture problem makes that an unsuitable model to copy wholesale. Someone should be able to save a thought, reopen a report and review an approval on a phone. The same objects should survive a layout change, even when a complex editor becomes full-screen. **Trade-off:** mobile supports sequential work rather than simultaneous evidence, chat and output panes.

### A proposed width budget, not a usability law

Assume a 240px navigation area, a 480px output editor and 32px combined spacing:

| Viewport | Chat space with all three areas | Chat space with navigation hidden |
|---|---:|---:|
| 1280px | 528px | 768px |
| 1366px | 614px | 854px |

This calculation explains the engineer's concern without claiming that 600px is a scientifically established failure threshold. Narrowing can increase wrapping, vertical scrolling and difficulty comparing tables, but **no reviewed evidence establishes a universal 600px cliff**. NN/g's information-processing account instead emphasises attention, working memory, screen capacity and prioritisation. [14]

**Implement this rule:** no persistent Actions column. At **1440px and above**, allow navigation plus chat plus an on-demand output editor. At **1200-1439px**, hide navigation while the editor is open. At **768-1199px**, switch the centre between **Chat** and **Work**. Below **768px**, use full-screen destinations with a clear Back action. These are proposed breakpoints; test them at browser zoom and in split-screen. **Trade-off:** less simultaneous visibility in exchange for usable content space and predictable navigation.

## 4. Output management: make the artefact independent, but keep its provenance

### How the major products handle durable work

| Product | Storage and return path | Naming, editing and version behaviour | Separation from chat |
|---|---|---|---|
| Claude | **Artifacts** in the sidebar collects work across conversations. [19] | Direct or conversational editing; editing an earlier message can create another chat version with its own artefacts. [19] | Separate retrieval area, linked creation context. |
| ChatGPT | **Library** contains uploaded and created files; search and origin/type filters aid return. [12] | Canvas supports history, restoration and **Show changes**. Do not assume every Library file has Canvas version behaviour. [10] | Deleting a chat does not delete its Library files. [12] |
| Gemini Canvas | Reopen the chat containing the document; edits are automatically saved. [11] | A shared Canvas can be copied and edited independently; a comprehensive version model is not established here. [11] | Primarily conversation-linked in the reviewed help. |
| Copilot Pages | A durable, editable page opens beside chat. [6] | Colleagues can access the page without accessing the originating chat. [6] | Explicit separation of shareable page and private exchange. |
| NotebookLM | Outputs and notes live within a notebook's Studio; notebooks are independent. [49] | Written notes are editable, but saved response notes are documented as non-editable; exported edits do not synchronise back. [7] | Project-like notebook container, distinct Sources and outputs. |
| Gamma | Gammas are organised in workspace folders. [33] | **Version History** restores earlier content; preserving the current alternative requires duplication before restore. [42] | Document-centred organisation rather than chat-only retrieval. |
| Perplexity | Projects consolidate sessions, files and tasks. [9] | **Create page** is temporarily retired; **Convert to Page** is described as returning. [3] | Do not design around Pages as an unquestioned stable pattern. |

The strongest common lesson is not "use folders". It is **give work a durable return path and make its relationship to conversations explicit**.

### Case study: Copilot Pages separates sharing from the conversation

Copilot lets a useful response become an editable page, and colleagues can work on that page without access to the original chat. That is a valuable distinction for enterprise research: an executive needs the deliverable, not necessarily the researcher's tentative reasoning or unrelated discussion. [6]

Compass should therefore give each asset a stable ID, title, project association, owner, type, status, timestamps and source links. Opening a report should not require locating the conversation that created it. Include **View originating conversation** as a secondary route. **Trade-off:** maintaining object-level permissions and lineage is more work than storing message attachments.

### Recommended object and version model

Use one **Library**, defaulted to the active project, with **All / Outputs / Sources / Ideas** views. Uploads and generated outputs can appear together in All, but carry unmistakable origin and type labels. A customer transcript must not be confused with an AI-generated summary of it.

Autosave edits to a stable asset; create recoverable revisions. **Regenerate** creates another revision after showing the intended scope; **Create alternative** creates a linked sibling. Never silently replace an approved report. Suggested titles should include project and subject, remain editable and not become identifiers. **Trade-off:** a small amount of version UI prevents both accidental loss and duplicate-file proliferation.

Make project association canonical provenance. Later, allow user-defined **Collections** as additional organisation, not as replacements for the originating project. A cross-project collection must not broaden access. For v1, defer arbitrary nested folders and place genuinely unassigned legacy assets in **Unassigned**, rather than guessing their project. **Trade-off:** less organisational freedom initially, but a simpler and safer migration.

## 5. Idea capture: save interpretation without turning it into evidence

### Patterns across the requested research and note-taking tools

| Tool | Documented pattern | What Compass should learn |
|---|---|---|
| Dovetail | Highlights become searchable clips and can be embedded in documents with raw-data connections; AI suggestions can be approved or rejected. [56] | Keep evidence links intact and suggestions reviewable. |
| EnjoyHQ | **Stories** combine findings with quotes, videos, reports, attachments and embedded highlights. [22] | An interpretation needs more expressive space than a quote bookmark. |
| Condens | Insights remain connected to sources; the user retains final control over AI output. [21] | Evidence-backed synthesis should remain inspectable. |
| Notably | Sticky notes can be clustered by theme, sentiment or journey stage; users can move them manually. [31] | Automation should reorganise thinking reversibly. |
| Marvin | Notes, transcripts, Insights and Playlists are searchable; project context narrows the search. [43] | Notes must participate in retrieval, not become a forgotten drawer. |
| Aurelius | Findings can retain supporting data; recommendations can be linked to insights; project reports are generated. [18] | Connect captured thought to subsequent action and reporting. |
| NotebookLM | **Add note** supports original thought; **Save to Note** captures a response; notes can become sources. [7] | Distinguish authored interpretation from imported evidence. |
| Obsidian | Daily notes offer a low-structure capture point; backlinks expose linked and unlinked mentions. [27][54] | Capture first, connect later. |
| Readwise | Highlights retain a **View in [Source]** route and can resurface through review and similarity search. [57] | Retrieval and resurfacing are part of the notebook, not optional decoration. |

These are documented capabilities, not comparative adoption rankings. Public evidence does not establish which clustering or resurfacing feature gets used most by enterprise researchers.

### Case study: EnjoyHQ separates a finding from its supporting material

EnjoyHQ's Story is not merely a saved quotation. It lets an author explain a finding and embed the material behind it, while collaborators comment and permissions distinguish viewing from editing. The resulting object is designed to transfer learning to other teams. [22]

Compass should make a smaller version of that object available during analysis: **Save idea** opens a lightweight editable note with the current project and optional selected evidence already attached. A standalone thought remains valid without a quote. Show **Hypothesis - not yet supported** when appropriate, rather than inventing a citation. **Trade-off:** ideas remain provisional and sometimes untidy, but researchers can think without completing a form.

### Capture now, structure when there is a retrieval need

The information-scraps study identifies lightweight entry, unconstrained content, flexibility, visibility and mobility as unmet needs. It does not provide a causal conversion-rate comparison between one-click capture and required tagging. [34]

**Recommendation:** require only the note text; infer the current project transparently; attach source IDs, transcript timestamps and the originating message when available. Save private to the author initially, with an explicit **Share with project** action. Put ideas in **Library > Ideas**, accessible from the chat toolbar, not in another top-level notebook application. **Trade-off:** one extra deliberate sharing action protects tentative or sensitive thinking.

For later use, add **Use selected ideas in a report**, search across ideas, and a project-level **Unresolved hypotheses** view. Defer automatic clustering and unsolicited digests until reuse is observed. Never let an AI-generated idea silently become primary research evidence: NotebookLM's note-to-source conversion demonstrates how easily authored material can re-enter a source pipeline. [7] **Trade-off:** fewer magical features, a cleaner distinction between observation, interpretation and evidence.

## 6. Saved prompts and routines: reuse a method before automating it

| Product | Reusable unit | Scheduling or control distinction |
|---|---|---|
| ChatGPT custom GPTs | Instructions, Knowledge, tools and Preview; management includes version history. [53] | The current help page says custom GPT retirement is planned; it does not justify treating the format as permanent. [53] |
| ChatGPT tasks | **Scheduled**, task instructions and a schedule; edit, pause or delete through a task panel. [25] | Actions requiring approval can pause execution. [25] |
| Claude Projects | Curated knowledge, chats and project instructions. [29] | A persistent working context is not itself a scheduled routine. |
| Gemini Gems | Named instructions with optional Knowledge files; access through **My Gems**. [38] | Reuses a configured method rather than implying recurring execution. |
| Perplexity Spaces/Projects | Current reviewed documentation uses **Projects** for ongoing searches, tasks, files and instructions. [9] | A container for ongoing work should not be conflated with a particular scheduling mechanism. |
| Zapier | A workflow can pause at **Request Approval**, including review or amendment of data. [28] | The approval step's test does not send reviewer notifications. [28] |
| n8n | Human review can be attached to a specific agent tool; reviewers see the tool and parameters. [39] | Approval executes the requested input; denial cancels the action. |

The important distinction is **saved method**, **saved scope** and **scheduled execution**. Those are different user promises and should not share an undifferentiated "saved prompts" list.

### Case study: n8n puts approval at the consequential boundary

n8n does not merely ask whether a user trusts the agent in general. Its documented human-review mechanism exposes a particular tool invocation and its parameters, then pauses until approval or denial. This makes the pending consequence inspectable. [39]

Compass should eventually describe a routine in plain language: "Every Monday, summarise new interviews in Project X, compare with the previous run, save a draft report, and notify the owner." Show its sources, time zone, next run, estimated resource use, owner and publication rule. Require **Run once** before scheduling, and show a result preview plus any skipped or inaccessible sources. **Trade-off:** configuration is longer, but non-technical users can understand what will happen without reading a workflow graph.

### Uptake evidence is weaker than agent enthusiasm

Qualtrics reports that **15%** of surveyed market research professionals actively use AI agents. That is not a scheduled-routine adoption rate and does not identify which job roles repeatedly run scheduled analysis. Its forecast that agents will handle much more work is an expectation, not current behaviour. [24]

No credible routine-specific B2B uptake benchmark was established in the reviewed material. Potential early adopters are continuous-listening and research-operations teams, but that is an inference to test, not a measured market segment.

**Decision:** keep team-approved actions in Actions; later put personal saved methods under **Actions > Saved** and scheduled executions under **Actions > Routines**. Store each run's results in Library with the relevant project. Do not build scheduling in v1. **Trade-off:** delayed convenience avoids stale permissions, recurring spend, duplicate reports and ownerless automations before repeat demand is proven.

## 7. Naming and information architecture: use labels that predict the destination

### Actual labels in documented production interfaces

| Product | Action or working surface | Durable-work destination |
|---|---|---|
| NotebookLM | **Studio**, **Chat**, **Sources** | **Notes**, **Reports**, **Slide Decks** and other item types within Studio. [49] |
| ChatGPT | **Canvas**, composer toolbox | **Library**, **Images**. [10][12] |
| Claude | **Artifacts**, templates | **Artifacts** in the sidebar. [19] |
| Gemini | **Canvas**, **Gems** | Conversation-linked Canvas; **My Gems** for reusable configurations. [11][38] |
| Copilot | **Prompt Gallery** | **Copilot Pages**. [46][6] |
| Dovetail | **Chat**, **Home** | **Browse**, folders, projects and docs. [44] |
| UserTesting | **Insights Discovery** | **Insights Hub**, reports and **Highlight reels**. [20][17] |
| EnjoyHQ | Story editor | **Stories**. [22] |

There is no converged industry name for "everything that is not chat". Copying a branded noun does not establish that customers will predict its contents.

### Case study: ChatGPT's Library combines a generic destination with concrete types

ChatGPT uses **Library** as a broad destination, then lets people distinguish uploaded from generated files and filter by familiar types such as documents, spreadsheets, presentations and PDFs. The generic container does not have to carry the entire information architecture. [12]

That is the model Compass should adopt: **Library** at the top level, then project context, origin and recognisable item types. Information-scent theory explains why the surrounding cues matter: people estimate the likelihood and cost of finding an answer using link labels, context and prior knowledge. It does not establish that "Library" universally beats "Reports" or "My Work". [58]

### Direct naming recommendations

Call the capability launcher **Actions**. Call the saved-work destination **Library**. Within a project, use **Outputs**, **Sources** and **Ideas** as views; in item lists, show precise types such as Report, Presentation and Showreel. Keep **Chats** separate. **Trade-off:** these labels are less distinctive than Studio, but make fewer promises users must decode.

Treat the following connotations as hypotheses, not research findings: **Lab** may imply experimentation; **Vault** may imply restricted or dormant storage; **Creations** may suggest a creative consumer tool; **Explorer** may suggest browsing rather than making. **My Documents** excludes video, shared work and ideas, while **My Work** can imply personal ownership of team assets. Validate these interpretations through tasks, not a naming popularity vote.

Use only two primary destinations, **Chats** and **Library**, with project switching available in both. Put Actions at the point of invocation and the Ideas view inside Library. This distinguishes the four object classes without four competing navigation systems. **Trade-off:** users sometimes switch a Library view, but the global navigation remains compact.

Do not split v1 actions into rigid Create and Analyse tabs. A report may be the result of analysis, and a persona may be both a model and an output. Use search and task labels first; introduce categories only if the customer study produces a stable grouping.

## 8. Enterprise market context: compete on trustworthy deliverables, not format count

### Market signals that matter to this decision

Qualtrics' global study covers more than **3,000 market research professionals across 14 countries**, fielded in Q3 2025. It reports movement towards purpose-built research AI: embedded AI use increased from **62% to 66%**, while general-purpose chatbot use declined from **75% to 67%**. These are vendor-reported survey findings, not proof that a particular interface caused a shift. [24]

The commercial implication is that Compass should make existing research work easier to complete and govern. A large format catalogue alone is not a durable differentiator.

| Insights platform | Concrete output/repository behaviour | Relevance to Compass |
|---|---|---|
| Dovetail | Project/folder navigation and cited answers; highlights embedded in docs connect findings to raw data. [44][56] | Searchable evidence and synthesis. |
| Qualtrics | Reports can be copied, renamed, filtered, exported to PDF and shared by link. [45] | Familiar report management and stakeholder distribution. |
| Remesh | **Export Data** offers participant CSV, conversation CSV and Topline PowerPoint; slides are generated per question. [50] | Data hand-off and editable executive decks. |
| UserTesting | Reports combine survey results, clips and findings; **Add to report** and multiplayer editing support assembly. [20] | Evidence-to-deliverable workflow. |
| Discuss.io | Project **Insights > Highlight Reels** supports presentation-ready video and role-based creation/viewing. [48] | Showreels are a substantive research job. |
| Canvs | XLSX analysis exports, a treemap PNG and comparison exports support external reporting. [52] | Analysts need editable data and presentation material. |
| Relative Insight | Its documented output proposition routes findings to teams and existing systems; detailed folder/version behaviour is not established here. [51] | Delivery into existing business workflows matters. |
| Notably | Research templates, tagging, clustering and source/ownership tracking support synthesis. [31] | Methods and provenance, not just finished formats. |

These examples support reports, presentations, themes, evidence and video clips as plausible priorities. They do **not** establish a global frequency ranking of deliverables. Obtain Compass's actual ranking from telemetry and recent customer work.

### Case study: Remesh and UserTesting support the last mile

Remesh's Topline PowerPoint produces editable slides per question, while UserTesting lets teams assemble findings, charts and clips in a report and share or present it. Neither documented workflow assumes the generated narrative is the end of the researcher's job. [50][20]

Compass should prioritise an editable report, an executive presentation and evidence-linked showreels. An interactive report may be highly valuable for some clients, but podcasts and infographics should not outrank these jobs merely because they make a stronger demonstration. **Trade-off:** less launch spectacle, more compatibility with existing delivery practice.

### Governance is part of output quality

ESOMAR's questions cover validation, data quality, lineage and human oversight. Ipsos found that generative AI could produce good topline summaries while struggling to elevate insights and business implications. These findings support a meaningful boundary between summarisation and recommendations. [60][59]

**Recommendation:** inherit project access by default; record who generated, edited, approved and shared an output; preserve source scope and prompt/model versions; distinguish draft from approved work; review external sharing separately. For global deployment, preserve source language, translation provenance, consent constraints, retention policy and local time zones. These are proposed enterprise requirements, not claims that every competitor supplies them. **Trade-off:** governance adds friction at approval and sharing, but should not interrupt every exploratory thought.

## 9. Mid-conversation actions: create a linked work thread with explicit scope

### Four interaction options

| Pattern | Documented analogue | Advantage | Accepted cost or risk |
|---|---|---|---|
| Inline continuation | Gemini permits transformations of a document reached through its existing chat. [11] | Maintains immediate conversational continuity. | Long outputs can bury the original discussion. |
| New conversation | Project containers retain related chats, files and instructions. [5] | Cleaner task boundary. | Context must be transferred explicitly; project membership alone is not a complete brief. |
| Branch | ChatGPT documents branched project chats appearing alongside the original. [5] | Preserves an alternative path. | More threads to manage. |
| Side-by-side work surface | Copilot Pages and Claude artefacts open beside chat. [6][19] | Separates editable work from conversational exchanges. | Requires space and can obscure which context governs edits. |

A side panel is a display decision, not a context policy. A branch is a conversation decision, not proof that every relevant source is included.

### Case study: Dovetail exposes scope while Marvin distinguishes search from synthesis

Dovetail lets users ask a grounded question and use an `@` mention to target a folder or project. Marvin distinguishes finding a named file from asking an analytical question and automatically constrains searches inside a project. Both make scope or retrieval mode visible rather than requiring users to trust an unspecified "whole workspace" operation. [44][43]

Compass should let a researcher say, "Compare recent interviews with lapsed customers in our UK pricing studies." It should resolve that into a compact scope card showing the matched projects, dates, segment rule and excluded material, then ask for confirmation. **Trade-off:** one explicit review step, but no requirement to operate a complex filter builder.

### The default rule ten messages into a chat

**Recommendation:** a substantial Create or Analyse action opens a **linked work thread** within the same project, with a breadcrumb to the originating message. The parent chat receives a small status/result card, not the full output. Preserve its scroll position and unfinished composer text. A selected-text transformation, such as shortening a paragraph, stays inline. If no project exists, ask the user to select or create one before generating a durable research asset. **Trade-off:** one additional task boundary prevents context contamination and lost place.

Pass a structured execution brief containing the explicit task, selected sources, project instructions, confirmed audience/segments/dates and relevant decisions from the conversation. Do not automatically pass every message or an arbitrary last-N window. Show the scope summary, not the model's hidden reasoning. Exact token-level retrieval policies are not disclosed consistently by the compared products, so they should not be imitated by guesswork.

For a large project set, use conversational disambiguation plus a readable selected-project list. Always allow removing an item and seeing all matches. Conversation-only selection can become slower and less inspectable than filters; retain optional direct controls as an escape route. **Trade-off:** slightly more UI, substantially better recoverability.

## 10. Measurement: measure discovered value and successful return

### Case study: Copilot benchmarks measure activity, not interface causality

Microsoft's published benchmark design compares active-user percentage, adoption by app and returning-user percentage across internal groups and peer organisations. That is useful for locating adoption gaps. It does not prove that a prompt gallery or any other interface component caused those gaps. [41]

Compass should use the same distinction. A button impression is exposure; an accepted output is value; a later return is retrieval success. A new Library may reduce chat activity while improving the product. **Trade-off:** a multi-stage funnel is harder to report than a single engagement number, but it avoids optimising the wrong behaviour.

### Proposed metric definitions

| Metric | Definition | Decision it supports |
|---|---|---|
| Capability activation | Eligible active users completing a first accepted result for an action / eligible active users | Did people discover something useful? |
| Time to first value | Time from first eligible opportunity to first accepted result | Is discovery or setup slowing use? |
| Capability breadth | Share of eligible active users completing at least three distinct action types in 28 days | Is Compass becoming broader than chat summarisation? |
| Opportunity-adjusted conversion | Accepted results / action impressions when prerequisites are satisfied | Does this button earn its prominence? |
| Repeat use | First-time completers who complete it again within 28 days / first-time completers | Was the initial use more than curiosity? |
| Output retrieval success | Users opening the correct existing asset without regeneration / retrieval attempts | Is the Library solving the stated problem? |
| Retrieval latency | Time from expressed retrieval intent to opening the right asset | Is finding work genuinely faster? |
| Idea reuse | Ideas later reopened, linked or incorporated into an output / saved ideas | Is capture creating useful memory rather than clutter? |
| Research quality | Human rubric score for fidelity, traceability, coverage and warranted implications | Is faster work still research-grade? |
| Guardrails | Wrong-project runs, permission failures, unsupported claims, unintended launches and overwritten approved work | Is convenience introducing unacceptable risk? |

These definitions are proposed, not published market norms. Set eligibility by job opportunity: a user without recordings should not count against showreel adoption.

### Instrumentation and decision rules

Log `action_impression`, `action_opened`, `scope_confirmed`, `generation_started`, `generation_completed`, `output_accepted`, `output_edited`, `output_shared`, `asset_reopened`, `idea_saved`, `idea_reused` and `action_abandoned`. Include action/version, entry point, project, source count, viewport class, latency and error category. Avoid copying sensitive transcript text into analytics.

**Proposed release gates:** at least **10 of 12** moderated-study participants retrieve the correct output within **30 seconds**, and at least **10 of 12** save and later find an idea without assistance. No participant should unintentionally launch research or share externally. These are directional usability gates, not statistical population estimates.

After four weeks and at least 50 eligible users, review a button for demotion if fewer than 5% complete it and there is no important low-frequency job justification. Consider promotion when at least 20% complete it and 30% of those repeat within 28 days. These are product governance rules, not significance thresholds. **Trade-off:** low-frequency strategic actions need qualitative judgement rather than automatic deletion.

Run discovery experiments separately from storage changes, randomising by workspace where collaboration could contaminate variants. Power the experiment from the observed baseline. Session-replay success means users stop scrolling old chats, reopening the wrong file, regenerating lost reports and keeping ideas in unrelated tools. **Trade-off:** careful observation costs time but explains why the funnel changed.

## 11. Customer research: a two-week study that tests behaviour, not enthusiasm

### Case study: tree testing answers a different question from card sorting

Card sorting asks participants to group and label items, exposing their mental model. Tree testing asks them to find a destination in a proposed hierarchy without the surrounding visual design. NN/g notes that tree testing can measure success and time, but generally does not explain why people behave as they do. [36][35]

Use the methods together rather than treating a survey asking "Do you like Studio?" as navigation validation. The moderated sessions explain ambiguity; the subsequent tree test checks whether the revised structure is findable. **Trade-off:** two small activities instead of one survey, with much stronger decision relevance.

### Exact proposed study

Recruit **12 participants from six customer accounts**: six hands-on researchers, three insights managers and three occasional collaborators. Recruit four in the Americas, four in EMEA and four in APAC, and include both frequent and light Compass users. This is a deliberately varied qualitative sample, not a representative global market survey.

| Activity | Sample and duration | What happens |
|---|---|---|
| Moderated interview and prototype test | 12 people, 35 minutes each | Eight minutes on recent work; 12 on tasks; seven on action grouping/priorities; five on naming; three on trust and controls. |
| Action-history questionnaire | Same participants, six minutes, after the unprimed tasks | Classify all 21 proposed jobs as used in Compass, done elsewhere, would try for a specified job, would not use, or unclear. |
| Unmoderated tree test | 30 additional users, 10 minutes each | Locate a report, transcript, idea, presentation, earlier version and action in the proposed structure. |
| Synthesis and decision | Product, design, research and engineering, 90 minutes | Decide labels, initial actions, recovery requirements and launch gates from observed failures. |

Run preparation on days 1-2, interviews on days 3-6, tree testing on days 7-9 and the decision session on day 10. **Trade-off:** fast directional evidence rather than a fully powered comparison of small differences.

### Questions and tasks

Ask: "Show me the last deliverable you sent to a stakeholder." "Where did the underlying evidence live?" "What did you save while analysing?" "When did you last fail to find something you had made?" "Which work did you repeat, and what changed between runs?"

Before showing the action catalogue, ask participants to create a report from supplied interviews, find a previously generated presentation, save an unsupported hypothesis and continue a task after ten chat messages. Counterbalance the folded-in and panel prototypes across participants; do not claim statistical superiority from this sample.

Then ask: "Which of these jobs have you actually done in the last month?" "Which would replace something you do now?" "Choose the three you would keep if the others disappeared." "What would you expect inside this destination?" "What must Compass confirm before proceeding?"

For naming, first collect spontaneous labels, then test **Library**, **My Work** and **Outputs** through destination tasks. Do not equate preference with findability. Avoid double-barrelled concepts, leading descriptions, marketing mock-ups, and treating "would use" as adoption. Record unfamiliar labels separately from rejected jobs. **Trade-off:** recent behaviour receives more weight than aspirational feature requests.

## 12. Risks and contrarian findings: the panel may be solving the wrong problem

### Case study: Perplexity Pages shows that shipped does not mean settled

Perplexity's help explicitly says **Create page** is temporarily retired and **Convert to Page** is intended to return with enhanced capabilities. This is a documented product change, but the source does not establish why it happened or prove that galleries failed. [3]

Compass should take the narrower lesson: do not use a competitor's current surface as validation of a durable customer job. Keep the content model, actions and layout separable so that a panel can change without moving or losing customers' work. **Trade-off:** additional architectural discipline now, less migration pain later.

### Failure modes to design against

| Risk | Evidence or status | Concrete response and accepted trade-off |
|---|---|---|
| Visible but still overlooked | Studio is prominent, yet a reviewer frames it as a "secret second app". This is anecdotal, not prevalence data. [32] | Add task-relevant examples and contextual entry points; accept some repeated exposure. |
| Confident summaries mistaken for insight | Ipsos reports weaker business implications than topline summarisation. [59] | Label generated recommendations as drafts requiring review; accept slower approval. |
| Export breaks the living document | NotebookLM says exported edits do not sync back and permissions do not carry over. [7] | Show export as a snapshot with its access consequences; accept a warning at export. |
| Regeneration damages trusted work | Gamma restoration replaces current content unless the user preserves a duplicate. [42] | Use recoverable versions and protect approved states; accept storage and UI complexity. |
| Cross-project synthesis leaks data | A plausible Compass risk, not a claimed incident. | Enforce source permissions at execution and sharing; accept incomplete answers when access is restricted. |
| Prompt buttons conceal methodology | Novices in the prompting study did not use systematic evaluation. [15] | Expose method, scope and evidence, with professional evaluation behind each action; accept release overhead. |
| Ideas become circular evidence | Notes can be converted back into sources in NotebookLM. [7] | Preserve an interpretation/evidence distinction; accept less automatic reuse. |
| A routine outlives its owner or permission | A foreseeable operational risk, not a measured adoption finding. | Require an owner, expiry/review date and pause-on-access-change behaviour; defer scheduling. |

The hidden cost users may resent is not simply screen space. It is having to decide what category their task belongs to, discover what an innocent button will do, and clean up the generated material afterwards. Those are hypotheses for the study, not established findings about all users.

### Why not 21 prominent buttons?

There is no evidence-based universal "six is right, 21 is wrong" rule. The choice-overload literature contains both positive and null effects and does not establish a single threshold. [16]

The stronger Compass argument is operational: 21 first-class actions create 21 promises about inputs, quality, context, recovery and ownership. Six is a proposed prioritisation decision, not an appeal to a magic number. Keep the rest discoverable in a directory when implemented, rather than pretending the assistant cannot do them.

Challenge four assumptions directly. Customers may have a **retrieval** problem rather than a desire for Studio. A showreel may be more important than a podcast even if both are easy to generate. Conversation-only project selection can be less efficient than a readable list. Internal staff discovering a feature late does not establish that all customers need it.

**Decision:** build reversible storage and capture improvements now, while researching the action hierarchy. Waiting for all research delays relief from known problems; skipping research risks building a large, attractive menu of low-value promises. **Trade-off:** accept some iteration, reject an irreversible all-at-once redesign.

## 13. Proposed v1: six prominent actions, one Library, one capture gesture

### Direct decisions on the open questions

| Open question | Recommendation | Trade-off accepted |
|---|---|---|
| Panel placement | Fold Actions into the composer; retain Library in the existing left navigation. No permanent Studio column. | Less ambient visibility than a full gallery. |
| Output editing | Open an on-demand editor; collapse navigation when space is constrained. | Occasional navigation switching. |
| Mobile | Full-screen Chat, Library and Work destinations with preserved state; idea capture and retrieval remain available. | Sequential rather than simultaneous work. |
| Mid-chat action | Linked work thread for substantial jobs; inline for selected-text edits. | More task objects, less conversational contamination. |
| Name of capability area | **Actions**. | Less brand personality. |
| Name of saved-work area | **Library**, with project and type views. | A generic label needs concrete content previews. |
| Prompt ownership | Product accountable; research approves method; design owns interaction; engineering owns implementation and evaluation. | Cross-functional review. |
| Folders | Project grouping now; user-defined Collections later. | Less initial flexibility. |
| Launch timing | Foundations and customer study in parallel; action prominence gated by findings. | Some reversible rework. |

This scope follows the documented separation between conversations and durable work in Copilot Pages and ChatGPT Library, without copying their exact interfaces. [6][12]

### Disposition of every proposed button

| Proposed action | v1 treatment | Rationale and accepted trade-off |
|---|---|---|
| Create Research | **Prominent action 1**; guided draft and explicit launch approval | Core job; accepts checkpoints before consequences. |
| Build a Presentation | **Prominent action 2** | Editable stakeholder deliverable; accepts template constraints. |
| Write a Report | **Prominent action 3** | Primary synthesis surface; accepts review before approval. |
| Build an Interactive Report | Directory entry with an example and existing invocation | Fixes the exact hidden-capability problem without a new wizard; accepts an extra selection step. |
| Build an Infographic | Directory entry for the existing capability | Preserves access without assuming frequent need. |
| Build a Chart | Contextual entry when structured data exists, plus directory listing | Prevents unsupported charting; accepts prerequisite checks. |
| Generate a Podcast | Directory entry for the existing capability | Available but not promoted over research delivery. |
| Create a Video Summary | Directory entry for the existing capability | Retains the format without another first-class flow. |
| Create a Showreel | **Prominent action 4**, enabled when recordings exist | Evidence-led storytelling; accepts clip review and consent controls. |
| Write a Discussion Guide | Step within Create Research; standalone route in the directory if already implemented | Avoids duplicate workflows; accepts one extra route for standalone use. |
| Executive Summary | Report preset, not a separate prominent button | One underlying synthesis flow; accepts less standalone visibility. |
| Pull Key Quotes | **Prominent action 5**, with transcript links | Frequent-job hypothesis tied to evidence inspection; accepts source checks. |
| Generate Recommendations | Optional report section, explicitly draft | Prevents false authority; accepts human review. |
| Theme Breakdown | **Prominent action 6**, labelled **Analyse themes** | Clear analytical job; accepts a reviewed thematic structure. |
| Sentiment Overview | No new dedicated v1 workflow | Requires validation across language, context and method; accepts delayed convenience. |
| Compare Segments | Next dedicated analytical action after v1; existing supported chat path remains available | Valuable but requires careful scope and denominator handling. |
| Build a Persona | Defer dedicated flow | Requires enough evidence and an explicit distinction from invented archetypes. |
| Draft an Email Summary | Report export/share preset | Avoids another overlapping generation promise. |
| Map the Journey | Defer dedicated flow | A framework-specific deliverable needs its own evidence contract. |
| Find Archetypes | Defer dedicated flow | Overlaps persona/segmentation work and risks unsupported generalisation. |
| Trend Over Time | Defer dedicated flow | Requires comparable dates, methods and source coverage. |

The six prominent actions are **Create research, Build a presentation, Write a report, Create a showreel, Pull key quotes and Analyse themes**. They are an initial hypothesis to test, not a claim that customer frequency has already been measured. The catalogue should list only implemented capabilities; do not expose speculative actions as working buttons.

### Candidate additions and minimum engineering work

Include editable Ideas, project association, Library search/type views, stable asset links, recoverable revisions, private drafts and explicit sharing. Add a capability description and example output to every directory entry. Migrate known project associations automatically; ask about unassigned assets rather than inferring ownership.

Defer scheduling, account-wide gap analysis, trend radar, topical digests, a separate persona builder and nested folders. They expand permission scope, evaluation demands or information architecture without being necessary to solve the three stated problems. Allow an existing chat request to remain the fallback where the capability is already supported. **Trade-off:** a deliberately less spectacular v1, but one that is easier to understand, evaluate and maintain.

## 14. Synthesis: the winning model is a research workspace, not a gallery

| Strategy and representative products | Mechanism | Scope and time horizon | Principal trade-off | Evidence strength |
|---|---|---|---|---|
| Source-grounded transformation: NotebookLM | Make source-to-output transformations explicit | A bounded notebook or project | Excellent coherence, potential cross-project and mobile limitations | Documented behaviour; no isolated panel-adoption effect. [49][8] |
| Conversation plus editable work: Claude, ChatGPT Canvas, Gemini Canvas, Copilot Pages | Turn an exchange into a persistent artefact | Immediate creation and subsequent editing | Better focus, but context, sharing and versions must stay intelligible | Strong workflow documentation; incomplete responsive specifications. [19][10][11][6] |
| Research repository: Dovetail, EnjoyHQ, Condens, Marvin, Aurelius, UserTesting | Connect evidence, interpretation and retrieval | Across studies and over time | More metadata and governance | Documented research workflows, limited comparative adoption evidence. [56][22][21][43][18][20] |
| Capture and resurfacing: Obsidian, Readwise, Notably | Preserve fragments, then connect or regroup them | From an immediate thought to later synthesis | Low-friction capture creates later organisation work | Descriptive features plus information-management research. [54][57][31][34] |
| Repeatable execution: Gems, project instructions, ChatGPT tasks, Zapier, n8n | Reuse instructions, then optionally automate timing and action | Repeated work and ongoing operations | Efficiency versus stale scope and unintended consequences | Strong control-pattern documentation, weak routine-specific uptake evidence. [38][29][25][28][39] |

The non-obvious tension is that **discovery benefits from visible possibilities, but trustworthy research benefits from narrower, explicit execution scope**. A large permanent gallery improves neither automatically. A visible catalogue can explain what Compass does; a scope-confirmed workflow determines what it does now.

A second tension is between frictionless capture and defensible synthesis. The same idea should be easy to save and hard to misrepresent as evidence. That requires a lightweight note interface on top of a rigorous object model, not a heavy form at the moment of inspiration.

Finally, artefact independence and project provenance are complementary, not competing. Users should return to a report without reconstructing a chat, while still being able to inspect the sources and decisions that produced it. The Library is the durable knowledge layer; chat is the interaction layer; Actions is the discovery layer.

**Final decision:** invest first in that separation. Keep the layout reversible, make the six visible workflows measurable, and let observed repeat use justify more automation. **Trade-off:** Compass gives up the immediate visual impact of a large Studio panel in exchange for a clearer path from question to evidence, deliverable and later reuse.

## References

URLs below identify the sources supporting the inline citations. Product documentation establishes behaviour, not causal adoption improvements.

### Assistant interfaces, outputs and reusable workflows

- Google, NotebookLM redesign: https://blog.google/innovation-and-ai/models-and-research/google-labs/notebooklm-new-features-december-2024
- Google, notebook creation and Studio outputs: https://support.google.com/notebooklm/answer/16206563?hl=en
- Google, NotebookLM mobile capabilities: https://support.google.com/notebooklm/answer/16296687
- Google, creating and adding notes: https://support.google.com/notebooklm/answer/16262519?hl=en
- OpenAI, Canvas: https://help.openai.com/en/articles/9930697-what-is-the-canvas-feature-in-chatgpt-and-how-do-i-use-it
- OpenAI, Library: https://help.openai.com/en/articles/20001052
- OpenAI, Projects: https://help.openai.com/en/articles/10169521-using-projects
- OpenAI, scheduled tasks: https://help.openai.com/en/articles/10291617-scheduled-tasks-in-chatgpt
- OpenAI, creating and editing GPTs: https://help.openai.com/en/articles/8554397-creating-a-gpt
- OpenAI, prompting guidance: https://platform.openai.com/docs/guides/prompting
- Anthropic, artefact behaviour: https://support.claude.com/en/articles/17153992-what-are-artifacts-and-how-do-i-use-them
- Anthropic, Projects: https://www.anthropic.com/news/projects
- Google, Gemini Canvas: https://support.google.com/gemini/answer/16047321?hl=en
- Google, Gems: https://support.google.com/gemini/answer/15146780?hl=en
- Microsoft, Copilot Pages: https://support.microsoft.com/en-us/microsoft-365-copilot/how-microsoft-365-copilot-pages-works
- Microsoft, Prompt Gallery: https://learn.microsoft.com/en-us/microsoft-365/copilot/copilot-prompt-gallery
- Microsoft, Copilot adoption benchmarks: https://techcommunity.microsoft.com/blog/viva_insights_blog/benchmarks-to-compare-copilot-adoption-coming-to-copilot-dashboard/4460432
- Perplexity, Projects/Spaces documentation: https://www.perplexity.ai/help-center/en/articles/10352961-what-are-spaces
- Perplexity, Pages status: https://www.perplexity.ai/help-center/en/articles/10352968-perplexity-pages
- Notion, AI entry points and capabilities: https://www.notion.com/help/guides/everything-you-can-do-with-notion-ai
- Cursor, reusable commands: https://cursor.com/changelog/1-6
- Cursor, Plan mode: https://cursor.com/help/ai-features/plan-mode
- Gamma, folders: https://developers.gamma.app/workspace/list-folders
- Gamma, version restoration: https://help.gamma.app/en/articles/11048579-can-i-undo-a-change-or-restore-a-previous-version-in-gamma
- Zapier, human approval: https://help.zapier.com/hc/en-us/articles/38731463206029-Request-approval-to-keep-your-workflow-running-with-Human-in-the-Loop
- n8n, human review of tools: https://docs.n8n.io/build/integrate-ai/ai-examples/human-in-the-loop-for-tools.md

### Research tools, market evidence and governance

- Dovetail, Home and project-scoped questions: https://docs.dovetail.com/help/home
- Dovetail, highlights: https://docs.dovetail.com/help/projects/highlights
- EnjoyHQ, Stories: https://help.usertesting.com/hc/en-us/articles/11880255202845-Stories
- UserTesting, Insights Hub: https://help.usertesting.com/hc/en-us/articles/24426406428317-Insights-Hub-Overview
- UserTesting, highlight reels: https://help.usertesting.com/hc/en-us/articles/11880419963549-Create-highlight-reels
- Condens, repository and evidence links: https://www.condens.io/
- Notably, synthesis workflow: https://www.notably.ai/
- Marvin, project/file/note search: https://help.heymarvin.com/en/articles/10130466-how-do-i-search-for-projects-files-and-notes
- Aurelius, research features: https://aureliuslab.com/features
- Qualtrics, report controls: https://www.qualtrics.com/support/survey-platform/reports-module/reports-section/reports-toolbar/
- Qualtrics, global market research trends findings: https://www.qualtrics.com/articles/news/research-teams-not-using-ai-are-four-times-more-likely-lose-organizational-influence
- Remesh, exports: https://help.remesh.ai/how-can-i-export-the-data
- Discuss.io, highlight reels: https://discussio.zendesk.com/hc/en-gb/articles/4625179144977-Highlight-Reels-for-Qualitative-Research
- Canvs, export formats: https://www.canvs.ai/resources/the-beginners-guide-to-canvs-exports
- Relative Insight, output proposition: http://www.relative.ai/
- ESOMAR, questions for buyers of AI research services: https://community.esomar.org/uploads/public/library/documents/miscellaneous/20-questions-to-help-buyers-of-ai-based-services.pdf
- Ipsos, AI quality in qualitative workflows: https://www.ipsos.com/en-us/conversations-ai-part-ii-unveiling-ai-quality-qualitative-workstreams

### Human-computer interaction, capture and evaluation

- Zamfirescu-Pereira and colleagues, *Why Johnny Can't Prompt*: https://dl.acm.org/doi/full/10.1145/3544548.3581388
- Wang and colleagues, task-supportive LLM interaction: https://arxiv.org/pdf/2402.06170
- PromptCharm: https://dl.acm.org/doi/full/10.1145/3613904.3642803
- Scheibehenne, Greifeneder and Todd, choice-overload meta-analysis: https://www.semanticscholar.org/paper/Can-There-Ever-Be-Too-Many-Options-A-Meta-Analytic-Scheibehenne-Greifeneder/07775d812c00a97aaa57b55596b39e26b8baaa0b
- *Information scraps*: https://dl.acm.org/doi/10.1145/1402256.1402263
- NN/g, scaling interfaces: https://www.nngroup.com/articles/scaling-user-interfaces
- NN/g, information scent: https://www.nngroup.com/articles/information-scent
- NN/g, card sorting: https://www.nngroup.com/articles/card-sorting-definition
- NN/g, tree testing: https://www.nngroup.com/articles/tree-testing/
- Obsidian, daily notes: https://obsidian.md/help/plugins/daily-notes
- Obsidian, backlinks: https://obsidian.md/help/plugins/backlinks
- Readwise, reviewing highlights: https://docs.readwise.io/readwise/docs/faqs/reviewing-highlights
- XDA, NotebookLM Studio commentary: https://www.xda-developers.com/notebooklm-studio-panel-secret-second-app

## What surprised me

1. **A visible panel does not necessarily make its capabilities feel discoverable.** The "secret second app" description of NotebookLM Studio challenges the assumption that placing tools on screen is sufficient. It is a reviewer observation, not a population finding. [32]
2. **The strongest competitor lesson is object independence, not three columns.** Copilot can share a page without sharing the originating chat; ChatGPT keeps Library files when their chat is deleted. [6][12]
3. **There is no defensible universal 600px chat threshold or ideal button count in the reviewed evidence.** The case for a compact v1 is prioritisation, screen budgeting and evaluability, not a magic-number usability rule. The choice-overload literature itself is conditional. [16]
4. **Some of the most attractive models have important mobile gaps.** NotebookLM's documented notes/report limitations and Canvas's stated mobile availability make literal imitation particularly risky for idea capture. [8][10]
5. **A note can become a research integrity problem.** Turning an interpretation into a source can create circular support unless its origin remains explicit. NotebookLM supports note-to-source conversion; the risk is an inference for Compass's design. [7]
6. **One-click generation may increase the need for professional ownership.** Novices in the prompting study stopped after local improvement and did not use systematic testing. Hiding the prompt removes neither that problem nor responsibility for the result. [15]
7. **Competitor patterns are less stable than the brief implies.** Perplexity temporarily retired Create page, while OpenAI's current GPT help describes a planned retirement. Neither establishes failure, but both argue against coupling Compass's durable research model to a fashionable interface. [3][53]
8. **The smallest credible solution is not a smaller Studio.** It is a discoverable action catalogue, a project-first Library and evidence-aware idea capture. The accepted trade-off is less spectacle in return for a clearer, safer research workflow.

## References

1. *Everything you can do with Notion AI*. https://www.notion.com/help/guides/everything-you-can-do-with-notion-ai
2. *Guidelines for Human-AI Interaction - Microsoft Research*. https://www.microsoft.com/en-us/research/project/guidelines-for-human-ai-interaction
3. *Perplexity Pages | Perplexity Help Center*. https://www.perplexity.ai/help-center/en/articles/10352968-perplexity-pages
4. *Prompting - OpenAI API*. https://platform.openai.com/docs/guides/prompting
5. *Projects in ChatGPT | OpenAI Help Center*. https://help.openai.com/en/articles/10169521-using-projects
6. *How Microsoft Copilot Pages works | Microsoft Support*. https://support.microsoft.com/en-us/microsoft-365-copilot/how-microsoft-365-copilot-pages-works
7. *Create & add notes in NotebookLM - NotebookLM Help*. https://support.google.com/notebooklm/answer/16262519?hl=en
8. *Get started with the NotebookLM mobile app - Android - NotebookLM Help*. https://support.google.com/notebooklm/answer/16296687
9. *What are Projects? | Perplexity Help Center*. https://www.perplexity.ai/help-center/en/articles/10352961-what-are-spaces
10. *What is the canvas feature in ChatGPT and how do I use it? | OpenAI Help Center*. https://help.openai.com/en/articles/9930697-what-is-the-canvas-feature-in-chatgpt-and-how-do-i-use-it
11. *Create docs, apps & more with Canvas - Computer - Gemini Apps Help*. https://support.google.com/gemini/answer/16047321?hl=en
12. *File storage and Library in ChatGPT | OpenAI Help Center*. https://help.openai.com/en/articles/20001052
13. *NotebookLM announces NotebookLM Plus and other new features*. https://blog.google/innovation-and-ai/models-and-research/google-labs/notebooklm-new-features-december-2024
14. *Scaling User Interfaces: An Information-Processing Approach  - NN/G*. https://www.nngroup.com/articles/scaling-user-interfaces
15. *Why Johnny Can’t Prompt: How Non-AI Experts Try (and Fail) to Design LLM Prompts | Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems*. https://dl.acm.org/doi/full/10.1145/3544548.3581388
16. [[PDF] Can There Ever Be Too Many Options? A Meta-Analytic Review of Choice Overload | Semantic Scholar](https://www.semanticscholar.org/paper/Can-There-Ever-Be-Too-Many-Options-A-Meta-Analytic-Scheibehenne-Greifeneder/07775d812c00a97aaa57b55596b39e26b8baaa0b)
17. *Create highlight reels – Knowledge Base Home*. https://help.usertesting.com/hc/en-us/articles/11880419963549-Create-highlight-reels
18. *Aurelius - The more powerful research repository and insights platform*. https://aureliuslab.com/features
19. *What are artifacts and how do I use them? | Claude Help Center*. https://support.claude.com/en/articles/17153992-what-are-artifacts-and-how-do-i-use-them
20. *Insights Hub: Overview – Knowledge Base Home*. https://help.usertesting.com/hc/en-us/articles/24426406428317-Insights-Hub-Overview
21. *Condens | UX Research Repository and AI Analysis Platform*. https://www.condens.io/
22. *Stories – Knowledge Base Home*. https://help.usertesting.com/hc/en-us/articles/11880255202845-Stories
23. *Slash commands, summarization, and improved Agent terminal · Cursor*. https://cursor.com/changelog/1-6
24. *Research Teams Not Using AI 4x More Likely to Lose Influence*. https://www.qualtrics.com/articles/news/research-teams-not-using-ai-are-four-times-more-likely-lose-organizational-influence
25. *Scheduled tasks in ChatGPT | OpenAI Help Center*. https://help.openai.com/en/articles/10291617-scheduled-tasks-in-chatgpt
26. *How do collaboration and sharing settings work in Gamma? | Gamma Help Center*. https://help.gamma.app/en/articles/11047226-how-do-collaboration-and-sharing-settings-work-in-gamma
27. *Daily notes - Obsidian Help*. https://obsidian.md/help/plugins/daily-notes
28. *Request approval to keep your workflow running with Human in the Loop – Zapier*. https://help.zapier.com/hc/en-us/articles/38731463206029-Request-approval-to-keep-your-workflow-running-with-Human-in-the-Loop
29. *Collaborate with Claude on Projects \ Anthropic*. https://www.anthropic.com/news/projects
30. *Task Supportive and Personalized Human-Large Language Model Interaction: A User Study*. https://arxiv.org/pdf/2402.06170
31. *Synthesis Platform for User Research | Notably*. https://www.notably.ai/
32. *NotebookLM's Studio panel feels like a secret second app - Here's how I make the most of it*. https://www.xda-developers.com/notebooklm-studio-panel-secret-second-app
33. *GET /folders*. https://developers.gamma.app/workspace/list-folders
34. *Information scraps: How and why information eludes our personal information management tools: ACM Transactions on Information Systems: Vol 26, No 4*. https://dl.acm.org/doi/10.1145/1402256.1402263
35. *Tree Testing: Evaluate Menu Labels and Categories - NN/G*. https://www.nngroup.com/articles/tree-testing/
36. *Card Sorting: Uncover Users' Mental Models - NN/G*. https://www.nngroup.com/articles/card-sorting-definition
37. *PromptCharm: Text-to-Image Generation through Multi-modal Prompting and Refinement | Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems*. https://dl.acm.org/doi/full/10.1145/3613904.3642803
38. *Use Gems in Gemini Apps - Computer - Gemini Apps Help*. https://support.google.com/gemini/answer/15146780?hl=en
39. *Human-in-the-loop for tools*. https://docs.n8n.io/build/integrate-ai/ai-examples/human-in-the-loop-for-tools.md
40. *How does the Readwise to Obsidian export integration work? - Readwise Docs*. https://docs.readwise.io/readwise/docs/exporting-highlights/obsidian
41. *Benchmarks to compare Copilot adoption coming to Copilot Dashboard | Microsoft Community Hub*. https://techcommunity.microsoft.com/blog/viva_insights_blog/benchmarks-to-compare-copilot-adoption-coming-to-copilot-dashboard/4460432
42. *Can I undo a change or restore a previous version in Gamma? | Gamma Help Center*. https://help.gamma.app/en/articles/11048579-can-i-undo-a-change-or-restore-a-previous-version-in-gamma
43. *How do I search for projects, files and notes? | Marvin's Help Center*. https://help.heymarvin.com/en/articles/10130466-how-do-i-search-for-projects-files-and-notes
44. *Home*. https://docs.dovetail.com/help/home
45. *Advanced-Reports Toolbar - Qualtrics*. https://www.qualtrics.com/support/survey-platform/reports-module/reports-section/reports-toolbar/
46. *Understand Prompt Gallery in Copilot | Microsoft Learn*. https://learn.microsoft.com/en-us/microsoft-365/copilot/copilot-prompt-gallery
47. *Research Repositories for Tracking UX Research and Growing Your ResearchOps - NN/G*. https://www.nngroup.com/articles/research-repositories/
48. *Highlight Reels for Qualitative Research – Discuss Help Center*. https://discussio.zendesk.com/hc/en-gb/articles/4625179144977-Highlight-Reels-for-Qualitative-Research
49. *Create a notebook in Gemini Notebook - Gemini Notebook Help*. https://support.google.com/notebooklm/answer/16206563?hl=en
50. *How can I export the data?*. https://help.remesh.ai/how-can-i-export-the-data
51. *Relative Insight | AI-Powered Customer Intelligence Platform : Relative Insight*. http://www.relative.ai/
52. *The Beginner's Guide to Canvs Exports | Canvs*. https://www.canvs.ai/resources/the-beginners-guide-to-canvs-exports
53. *Creating and editing GPTs | OpenAI Help Center*. https://help.openai.com/en/articles/8554397-creating-a-gpt
54. *Backlinks - Obsidian Help*. https://obsidian.md/help/plugins/backlinks
55. *Plan mode*. https://cursor.com/help/ai-features/plan-mode
56. *Highlights*. https://docs.dovetail.com/help/projects/highlights
57. *Reviewing Your Highlights - Readwise Docs*. https://docs.readwise.io/readwise/docs/faqs/reviewing-highlights
58. *Information Scent: How Users Decide Where to Go Next - NN/G*. https://www.nngroup.com/articles/information-scent
59. *Conversations with AI Part II: Unveiling AI quality in qualitative workstreams | Ipsos*. https://www.ipsos.com/en-us/conversations-ai-part-ii-unveiling-ai-quality-qualitative-workstreams
60. *ESOMAR 20 Questions to Help Buyers of AI-Based Services for Market Research and Insights*. https://community.esomar.org/uploads/public/library/documents/miscellaneous/20-questions-to-help-buyers-of-ai-based-services.pdf