PMind turns messy customer feedback into a ranked, evidence-backed backlog — automatically.
Most "AI for PMs" tools are a chat box that writes a nicer first draft. PMind's actual edge is upstream of that: a background multi-agent pipeline that reads your raw research (interviews, support tickets, NPS dumps) and turns it into severity-tagged insights, clusters them into themes, and scores opportunities with RICE — all cited back to the original customer quotes. The doc editor with Cmd+K is what you use after you already know what to build.
1. Ingest 2. Extract insights 3. Rank opportunities 4. Ship
───────────── ──────────────────── ─────────────────── ──────────────
Interview transcripts Background agents pull Opportunity agent clusters Commit an
Support tickets verbatim pain-point quotes, insights across themes, opportunity →
NPS / CSV exports ──▶ tag persona + sentiment, ──▶ scores Reach/Impact/ break into
assign severity 1–5, Confidence/Effort, and epics + tickets
sort into theme folders ranks by RICE score → sync to
Jira/Linear
- Automated insight extraction — drop in interview transcripts or support exports; background agents harvest verbatim quotes, tag the persona and sentiment, and assign a severity score, no manual tagging.
- Theme clustering — insights are grouped into reusable theme folders so you can see what customers are actually complaining about in aggregate, not one interview at a time.
- RICE-scored opportunities — the Opportunity agent clusters insights across themes into proposed opportunities, each scored on Reach/Impact/Confidence/Effort and traceable back to the exact quotes behind it.
- Decision ledger + outcome capture — every shortlist/commit/discard decision is recorded, and shipped features get revisited later against real outcomes, so the system learns which calls actually worked.
- Multi-agent handoffs — PM, Analyst, Designer, Opportunity, Calendar, and Whiteboard agents hand off to each other and synthesize findings, so asks like "synthesize my interviews, pull the perf numbers, tell me what to do" happen in one turn instead of five tools.
Once you know what to build, PMind also gives you an AI-native document editor: press Cmd+K anywhere to generate PRDs, ticket breakdowns, briefs, or stakeholder updates, grounded in a Product Brain (your product strategy/context, injected into every AI call) instead of generic templates.
| Cmd+K command picker | AI streaming into the doc | Result applied inline |
|---|---|---|
 |
 |
 |
▶️ Watch the demo GIF
- AI Chat sidebar — threaded, persisted chat with model picker and RAG over your knowledge base.
- Knowledge base (RAG) — research docs and interview notes are chunked, embedded (pgvector), and retrieved during chat, search, and discovery ingestion.
- Ticket export — generate structured tickets and push them to Jira or Linear.
- Projects & file tree — organize docs in folders per project, with debounced auto-save and global semantic + text search.
- Pluggable LLM providers — Gemini (default), Anthropic Claude, or OpenAI. Switch with a single env var; users can also override the model per request from the UI.
- Templates, theming, billing — PM document templates, light/dark mode, and optional Stripe subscription management.
pm_cursor/
├── cursor-for-pms/ # Next.js 15 frontend (port 3000)
│ └── src/
│ ├── app/ # App Router pages (projects, editor, billing, blog…)
│ ├── components/ # Editor, AICommandModal, CursorChat, Sidebar, …
│ └── store/ # Zustand stores (Product Brain, active project, editor)
├── backend/ # FastAPI backend (port 8000)
│ ├── main.py # App entry, CORS, router mounts
│ ├── prompts.py # System prompt templates per command
│ ├── llm/ # Provider abstraction (Gemini / Claude / OpenAI)
│ ├── agent/ # Multi-agent orchestrator, tools, specialist agents
│ ├── routers/ # ai, projects, documents, knowledge, integrations, billing, …
│ └── migrations/ # Discovery + longitudinal-memory SQL
├── supabase_*.sql # Core schema migrations (run in Supabase SQL editor)
├── landing/ # Static marketing page
└── testing/ # Fictional sample interviews to try research synthesis
| Layer | Choice |
|---|---|
| Frontend | Next.js 15 (App Router, TypeScript, React 19) |
| Editor | Tiptap (ProseMirror) |
| Styling | Tailwind CSS + shadcn/ui primitives |
| Auth | Clerk |
| Client state | Zustand |
| Backend | FastAPI + Uvicorn |
| LLM | Pluggable — Gemini (default), Claude, OpenAI, streamed over SSE |
| Database | Supabase (PostgreSQL + pgvector) |
| Integrations | Jira, Linear, Stripe (optional) |
All AI traffic flows through the FastAPI backend (/ai/*), which builds the system prompt from the command template + Product Brain context and streams the provider response back as Server-Sent Events.
- Node.js 18+
- Python 3.10+
- A Supabase project (database)
- A Clerk application (auth)
- An API key for at least one LLM provider: Gemini, Anthropic, or OpenAI
Run the SQL migrations in your Supabase SQL editor, in order:
supabase_schema.sql
supabase_phase0.sql
supabase_phase1_filetree.sql
supabase_phase2_chat.sql
supabase_phase2_rag.sql
supabase_phase2b_storage.sql
supabase_phase3_integrations.sql
supabase_phase4_billing.sql
supabase_phase4_search.sql
backend/migrations/document_chunks.sql
backend/migrations/discovery.sql
backend/migrations/tier1_longitudinal.sql
backend/migrations/tier2_decision_ledger.sql
backend/migrations/tier3_outcome_capture.sql
cd backend
python -m venv venv
venv\Scripts\activate # Windows
# source venv/bin/activate # macOS/Linux
pip install -r requirements.txt
cp .env.example .env # then fill in your keys
uvicorn main:app --reload --port 8000Switching LLM providers is just env vars — no code changes:
LLM_PROVIDER=gemini LLM_MODEL=gemini-2.5-flash GOOGLE_API_KEY=...
LLM_PROVIDER=claude LLM_MODEL=claude-sonnet-4-6 ANTHROPIC_API_KEY=...
LLM_PROVIDER=openai LLM_MODEL=gpt-4o OPENAI_API_KEY=...cd cursor-for-pms
npm install
cp .env.local.example .env.local # then fill in your keys
npm run devOpen http://localhost:3000, sign in, create a project, and press Cmd+K (or Ctrl+K).
The testing/ folder contains five fictional user interviews and research notes for a made-up product ("Shopflow"). Upload them as knowledge documents and open a project's Discovery tab to watch the full pipeline run: insight extraction with severity/persona tags → theme clustering → RICE-ranked opportunities. You can also paste one into a doc and run the interview synthesis Cmd+K command for a lighter-weight, single-doc version of the same idea.
Secrets live only in backend/.env and cursor-for-pms/.env.local — both are gitignored. Use the committed example files as templates:
backend/.env.example— LLM provider keys, Supabase URL + service key, Clerk secret keycursor-for-pms/.env.local.example— Supabase URL + anon key, Clerk publishable/secret keys, backend URL
Never commit real keys. The Supabase service key bypasses RLS and must stay server-side.
Issues and pull requests are welcome. Useful maps of the codebase:
CODEBASE_CONTEXT.md— file-by-file reference of the whole repoDESIGN.md— the UI design systembackend/agent/PLANS.md— the multi-agent orchestration roadmap
Before opening a PR, run npm run typecheck && npm run lint in cursor-for-pms/ and pytest in backend/.
MIT © 2026 Adamya Vashisth