Deep Research guide

What a run is

A run has five steps: Plan, Collect, Analyse, Verify and Report. You see the step you are on, and the sources move along in their own lanes while it collects.

1. 1

   Plan

   Type a product name and press Find it. Confirm one match per store, add the Reddit communities to read, and check the plan table. It lists each source, what it fetches, the requests and the time, and the number of calls to your AI.

2. 2

   Approve

   Press Approve and start. A plan above 200 requests asks you to confirm again. Switch a source off in the plan to skip it.

3. 3

   Collect

   Desktop reads the sources one polite request at a time. You can keep working, close the window, or press Finish with what we have once enough rows are in.

4. 4

   Analyse with your own AI

   Your AI agent reads the rows in batches and proposes findings in eight categories, each with quotes. The app saves only what it can check.

5. 5

   Report

   The app counts the findings, groups them into themes you can rename, merge or drop, and shows the report, the sources and a trace of what happened.

On the empty screen, press See an example. It is a made-up app with generated reviews, so you can look at a finished report without fetching anything.

Sources and their limits

• App Store and Google Play: reviews of the product you confirm. You check the match before anything is fetched.
• Reddit: needs at least one community, because Reddit history only searches inside a community. Desktop suggests communities from the Reddit archive by name, and, when Clipper is paired, from a live Reddit search in your own browser (Clipper 1.5.1 or later). Each community is checked and marked active, quiet or not found. History needs no login. Live Reddit only runs through a paired Clipper, and signing in to reddit.com in that Chrome gives better results.
• Hacker News: discussions that mention the product.
• Your sites: pages you saved from any website with Other sites (see Desktop research sources). Nothing is fetched in the run: it reads the saved pages that mention the product, quotes them word for word and names the site as the source. Off unless you switch it on for the run.
• Review sites (G2, Capterra, TrustRadius, Trustpilot): read through your own paired Clipper (1.5 or later) in a background tab, at a slow pace with a daily limit. Make a free account on each site and stay signed in to it in the Chrome where Clipper runs: signed-in visits get far fewer human checks and better results. Each site is off until you switch it on in Settings, Connect Clipper, after a short notice. You pick the product on each site first, then the lowest rated reviews are collected. If a page does not match the app's reader, your own AI reads it (see the Software Reviews Lab), and the plan shows "N pages are waiting for your AI to read". A review is kept only if its text is on the page word for word. If a site shows a check or asks you to log in, the run stops on that site and asks you to deal with it in your browser. Clipper never tries to get past a check.

Depth

Quick is the default. It reads up to 600 rows and is the fastest way to see whether there is something here. Standard reads up to 2,000 rows, so it takes longer to collect, and the composer shows about how many calls it needs from your AI before you start. If a source has more rows than the depth allows, Desktop takes a repeatable sample and the report says so.

The AI calls run on your own AI plan (Claude Code, Codex, OpenCode or Ollama). BigIdeasDB charges nothing per run, and there are no credits to buy. Sources are paced, so a larger run takes minutes, not seconds.

Read the report

• Summary and coverage: one chip per source with its rows and any flag, such as Partial or Blocked.
• Eight questions: each category shows how many analysed rows mention it, with quotes and an Open original link.
• Themes: groups of findings. Rename, merge or drop them and the app recounts.
• Sources and Trace tabs: rows per source and lane outcome, and what the app and your AI did.
• Partial banner: shown when a source was blocked, a cap cut it short, or you finished early.

A competitor in the results is validation, not a reason to stop. Look at what people still complain about, because that is where a wedge may be.

Use it from Claude Desktop

Connect Claude Desktop, Cursor, Codex or another MCP client in Settings, Connect your AI (see Connect your AI). Then run the built-in deep-research skill, or ask in plain words: "Run Deep Research on Headspace." The skill walks your AI through the run.

• Your AI creates the run, finds the stores and communities, and shows you the plan in plain words. It does not start without your yes.
• An AI client can start a plan of up to 200 requests. A bigger plan is approved in the Desktop window.
• Collecting takes minutes. Your AI tells you and ends the turn. Ask it for the run later and it carries on from where the run is.
• Your AI reads rows in batches and saves findings with quotes copied word for word. It sends no counts. The app counts.
• Your AI writes a short summary from the app's numbers and reports from the saved report only.
• The Deep Research group in the tools reference lists every tool with an example prompt.

Export and pin

• Export Markdown saves the report to a file on your computer.
• Pin a finding, or everything, to the evidence board as notes with their checked quotes and counts. Pinning twice keeps one card.
• Make idea turns a finding into an idea card, with the count as its demand proxy.
• Delete a run any time. The reviews and posts you collected stay in the project.

Nothing is uploaded. Collecting, counting, checking and exporting all happen on your computer, and only your own AI agent sees the rows.

If something goes wrong

• Start is off: confirm a product match and, for Reddit, at least one community.
• A source asked us to slow down: collecting continues more slowly, on purpose.
• A source was blocked or did not finish: press Resume. Windows already done today are skipped.
• The analysis stopped: check that your AI agent is signed in in Settings, then press Retry analysis. The rows you collected are kept.