Kit de Recherche AI pour Product Managers

What's in this pack

This is the stack for a PM who already has too many tabs open — not a thought-leader's "AI for product" reading list. Every pick here does one job in a real product week: open a discovery topic, capture the user interview, scan the four competitors your CEO just mentioned, draft a PRD that engineering doesn't bounce, get a roadmap view stakeholders can actually read, and watch the metrics so you know whether last sprint mattered.

The pack is agent-assisted on purpose, not autonomous. PMs don't get fired for being slow; they get fired for shipping the wrong thing. So the workflow keeps you in the loop at every decision point — AI handles the bulk research, transcription, and first drafts; you read, push back, and own the call.

No "learn LangGraph for two weeks first" — most of these install as a Claude Code subagent or a self-hosted app. Pick a discovery agent, pick a meeting tool, pick a PRD agent, and you can use the whole pack on Monday.

Install in this order

1. STORM — AI Research Report Generator (Stanford) — start here, because nothing else matters if you're researching the wrong topic. Feed it "how are PMs at fintech startups handling KYC onboarding" or "what do users complain about in " and you get back a structured, cited report. Stanford-built, multi-perspective conversation simulation — closer to a research analyst than a search box.
2. Claude Code Agent: Research Analyst — STORM gives you a report; this agent synthesizes across multiple reports, interview notes, and competitor scans into the patterns you actually need for a PRD. Pair them: STORM does the broad sweep, Research Analyst does the "so what" pass.
3. Meetily — Privacy-First AI Meeting Assistant — the moment user interviews matter, your laptop becomes a court reporter or your insight pipeline dies. Meetily runs locally with Whisper + Ollama. No cloud, no "please record this meeting" awkwardness for sensitive enterprise calls. Real-time transcription + speaker diarization + local summarization.
4. Anarlog — Open-Source AI Meeting Notes That Stay on Your Machine — alternative meeting recorder if Meetily doesn't fit your OS or workflow. Same posture: on-device only, your interview transcripts never leave the laptop. Pick one of the two and standardize.
5. Claude Code Agent: Competitive Intelligence Analyst — competitor scanning is the most over-promised, under-delivered AI task. This agent does it the right way: pulls public signals (pricing pages, changelog, blog, jobs, reviews), builds a comparison matrix, and flags the gaps that matter to your positioning. Re-run weekly so you don't get blindsided in a sales review.
6. Claude Code Agent: Product Manager — PRDs & User Stories — feed it the research from steps 1-2, the interview takeaways from steps 3-4, and the competitive matrix from step 5, and it drafts a PRD with user stories and acceptance criteria. First draft only — your job is to push back on the parts that hallucinate scope or skip edge cases. Saves ~3 hours per PRD; doesn't replace the thinking.
7. Claude Code Agent: Product Manager — different agent, different job. This one focuses on the prioritization call: "given these 12 features and this team capacity, what should the next sprint be?" Use it as a sparring partner before you walk into the prioritization meeting, not as the decision-maker.
8. Claude Code Agent: Product Strategist — for the larger move: positioning, GTM, the "should we even be in this market" question. Less useful for sprint planning, very useful when your CEO asks for a strategy doc by Friday.
9. Plane — Open-Source AI Project Management — issues, sprints, modules, roadmaps, real-time analytics. Self-hosted Jira/Linear alternative with AI-powered pages. 47,500+ stars. Use it as the single source of truth for the roadmap your PRDs ladder into. AGPL-3.0 means no vendor lock-in for the most political artifact you own.
10. PostHog — Open Source Product Analytics & Session Replay — the loop closes here. You shipped the thing the PRD described; PostHog tells you whether anyone used it, whether they got stuck, and whether the metric you promised actually moved. Session replay is the single most useful PM tool for catching the "the feature works but no one can find it" failure mode.

How they fit together

        ┌─── STORM ──── Research Analyst ───┐
        │  (broad sweep)   (so-what pass)    │
        └────────────────┬───────────────────┘
                         │
        ┌─── Meetily ─── or ─── Anarlog ────┐
        │       (user interview capture)     │
        └────────────────┬───────────────────┘
                         ▼
               Competitive Intelligence Analyst
                  (weekly competitor scan)
                         │
                         ▼
    ┌──── PM Agent: PRDs ──── PM Agent: Prioritization ────┐
    │     (draft PRD)         (sparring partner)            │
    └──────────────────────┬───────────────────────────────┘
                           │
                           ▼
                  Product Strategist
              (the larger positioning call)
                           │
                           ▼
                        Plane
              (roadmap → sprint → ticket)
                           │
                           ▼
                       PostHog
          (did anyone use it? did the metric move?)

The critical join is research → PRD → roadmap → analytics. Each handoff is where PMs lose the thread: the research never makes it into the PRD, the PRD never makes it into the sprint, the sprint never gets measured. These ten tools are picked so you can walk the chain end-to-end in one tool surface.

Tradeoffs you'll hit

• STORM vs just asking Claude to research a topic — Claude with web search is faster for a single question. STORM earns its place when you need (a) a cited report you'll attach to the PRD, (b) the same research repeated across 5 competitors or 8 personas in a batch, or (c) the multi-perspective expert simulation that surfaces angles a single prompt misses. For PM research, the citation trail matters — engineering and design will challenge claims.
• Meetily vs Otter/Fireflies/Granola — Meetily is on-device. The cloud options are slicker and have better summaries today, but for B2B PM you'll eventually interview a customer who can't legally have their voice on a third-party server. Run Meetily in parallel from day one so you're not scrambling when that interview gets booked.
• PM Agent: PRDs vs your existing PRD template — the agent doesn't replace the template; it fills it faster. If you have a battle-tested template, paste it as context and the agent will conform. If you don't have one, use the agent's default and start a template. Don't let the AI design your PRD format — that's a taste decision.
• Product Strategist vs Product Manager agents — Strategist thinks in quarters and markets; PM thinks in sprints and features. Both useful, neither replaces the other. Most weeks you live in the PM agents; once a quarter you escalate to Strategist for the board deck.
• Plane (self-hosted) vs Linear/Jira (SaaS) — Plane is open-source, self-hostable, AGPL-3.0. Linear is faster and prettier. Jira is what your enterprise already runs. Pick Plane if you control the stack and want to own your roadmap data; pick the others if the org has already standardized. The pack lists Plane because it's the option you can evaluate without filling out a procurement form.
• PostHog vs Mixpanel/Amplitude — PostHog is open-source and self-hostable; the SaaS competitors are more polished. For a PM kit, PostHog wins because (a) session replay is included not a $40/seat upsell, (b) feature flags + analytics in one tool means you can ship gradual rollouts without two integrations, (c) you can run it on your own data when the legal team asks.

Common pitfalls

• Over-trusting the AI-synthesized interview summary — the model will smooth over the quote where the user actually contradicted your hypothesis. Always re-listen to the 30 seconds before and after any quote the agent calls out as a "theme." The contradictions are where the insight lives.
• Vibe-based roadmap from competitive scans — the competitive analyst gives you a comparison matrix. That's an input, not a roadmap. "Competitor X has feature Y" is not a reason to build feature Y; it's a reason to ask why your users haven't asked for it. Use the matrix to find your wedge, not to copy.
• Letting a PRD agent ship without an engineering read — the PRD agent will confidently write "the system should handle 10x traffic" without knowing whether your backend can. Always do an engineering pass before the spec leaves your hands; otherwise the bounce comes after the agent's been gone three days.
• Confusing transcription with synthesis — a 90-minute interview transcribed verbatim is 12,000 words you'll never re-read. Meetily/Anarlog give you the transcript; the synthesis (themes, quotes, action items) is a separate AI pass. Don't skip it. The transcript is evidence; the synthesis is the artifact you actually use.
• Building the agent stack before talking to a customer — the worst PM failure mode is research-procrastination dressed up as tooling. Pick three tools (STORM + Meetily + PM Agent: PRDs), book a customer interview for tomorrow, ship a PRD by Friday. The other seven tools earn their slot only when the first three are saturated.