deep-research
A sourced market, user and technology study on one product question, with the key figures checked.
A polished market report with invented numbers is worse than none. This one opens the sources behind the figures its conclusions lean on, and says which research path actually ran.
- 2026
- Claude Code
- Web search, Gemini Deep Research
- MIT
View on GitHub (opens in a new tab)
git clone https://github.com/VladPolyakov-art/deep-research ~/.claude/skills/deep-researchWhat it writes
Fixed file names, so the other skills in the kit find them without asking. Each file also gets a row in research/INDEX.md.
- research/deep-research.md
- The verified report: findings per question, conflicts between sources, what it means for the product.
- research/deep-research-prompt.md
- The research plan, also a paste-ready prompt for Gemini.
- research/deep-research-raw.md
- Raw material from Gemini or sub-agents, unedited.
How it works
- Questions about your product5 to 7 of them, aimed at the gaps of the competitor scan and the personas, if you have them.
- Three ways to gatherClaude researches with web tools, or writes a prompt you run in Gemini Deep Research, or uses a delegate.py route if your project has one.
- Key figures checked3 to 5 claims the conclusions depend on are matched against opened sources. A pasted Gemini claim stays unverified until then.
What the agent does
- ContextReads the brief and sibling files. With nothing to go on, asks 3 to 5 questions.
- PathPicks how to gather the material and says which one.
- Plan5 to 7 questions, saved as a prompt that works on its own.
- GatherItself, through parallel sub-agents, or from the report you paste back.
- VerifyOpens the sources behind the key figures, fixes or drops what does not match.
- SynthesiseFindings, changes versus the earlier research, and what it means for the product.
Works with the kit
None of these is required. With an empty research/ folder it works from what you give it, and when a sibling’s file is there, it reuses it instead of redoing the work.
Reads
- competitive-researchThe competitor list and, above all, its gaps.
- proto-personasSegments, and the assumptions worth testing.
- ux-auditPains to test against public evidence.
- simulated-interviewHypotheses to check, still marked synthetic.
- feature-researchFeature requests worth checking against public evidence.
Read by
- feature-researchMarket context, instead of redoing it.
- proto-personasSegments, behaviours and voice of users.
- simulated-interviewGaps and open questions, as the riskiest assumptions.
- competitive-researchMarket facts and competitor names.
Part of the Product research kit: six skills, one research/ folder and one set of evidence tags.