Open source · MIT · Runs inside your AI coding agent
n2b turns a raw product idea into an investment-ready blueprint — market research, feature specs, architecture, and a handoff package any dev team or AI build tool can run with. It deliberately writes no code.
.n2b/ ├─ BRIEF.md · validated project brief ├─ features/ · research, personas, journeys, feature list ├─ specifications/ · one spec per feature, verbatim ACs ├─ architecture/ · recommended stack + DB schema └─ exports/ · jira · speckit · lovable-pack · …
Blueprint complete — ready to hand off
Where n2b fits
Your idea goes in rough. n2b runs the planning half of the project — research, definition, specs, architecture — and hands a finished blueprint to whatever builds it: an AI app builder, a coding agent, a tracker, or a team of humans.
you keep your build tools — n2b just goes first
Diagram: a raw idea flows into n2b. Inside n2b, stages 1 Intake, 2 Define, 3 Specify and 4 Architect each add to the blueprint; when stage 4 passes, the blueprint is complete. Stage 5 Export then renders it for AI app builders (Lovable, v0, Bolt, Replit), AI coding agents (Cursor, Codex, Devin), spec-driven kits (Spec Kit, Task Master, BMAD), issue trackers (Jira, backlog.json), or a human dev team. On the project timeline, n2b is the Plan phase, between Idea and Build.
Lovable, v0, Bolt, and Cursor build fast — from whatever you give them. Run n2b first and hand them a blueprint instead of a vibe: researched features, real acceptance criteria, a data model. Stage 5 even exports a per-feature prompt pack for your tool of choice.
See the prompt-pack exports →No PM, no architect, no researcher? The pipeline plays those roles — interviewer, market researcher, spec writer, architect. One interview in → a full feature list, per-feature specs, recommended architecture, and a database schema, before dev kickoff.
See what the blueprint contains →AI build tools get you 70% of the way in an afternoon. Then the skipped groundwork surfaces: no research, no feature definition, no data model, no acceptance criteria — and the build starts to drift. The fix isn’t better prompting. It’s doing the front half of the lifecycle before the first line of code.
A structured interview about your idea — it asks until it’s confident, not until a form is filled.
→ BRIEF.mdMarket research, personas, journeys, a prioritized feature list — audited for completeness.
→ features/Every feature in implementation-ready depth, with verbatim acceptance criteria and a hard quality gate.
→ specifications/Live tech research → a recommended architecture with documented alternatives, plus a database schema.
→ architecture/The finished blueprint, rendered for whichever team or tool builds it. Repeatable, fidelity-gated.
→ exports/Each stage passes a quality gate before the next will run. When stage 4 passes:blueprint complete.
Invoked as /n2b:s1-init in Claude Code · /n2b-s1-init in Cursor and OpenCode · $n2b-s1-init in Codex
Works where you already work
n2b installs project-locally into whichever AI coding agent you already pay for. No new account, no metered credits, no SaaS in the middle — the same pipeline, translated to each runtime’s command format.
npx napkin-to-blueprint --claude/n2b:s1-initnpx napkin-to-blueprint --cursor/n2b-s1-initnpx napkin-to-blueprint --opencode/n2b-s1-initnpx napkin-to-blueprint --codex$n2b-s1-initRun the installer with no flag and a picker asks which runtime to install; the --all flag installs every runtime.
Every export is checked against the blueprint by a fidelity gate before it’s accepted — counts reconciled, acceptance criteria carried verbatim.
See them in real runs ↓Showcase · real runs
Each idea went through every n2b stage, gates and retries included, and was published exactly as the pipeline wrote it. Open a run, read what it decided, then build it with the pack it ships.
“They lose real money to no-shows and last-minute cancellations…”
Looks like a calendar plus a form. The blueprint found the money flow a first draft misses: deposit, balance, no-show forfeit, refund window. Plus two-way calendar sync and no double-booked slots.
“…“AI meal plan” toys that suggest a walnut salad to a family with a nut allergy.”
The most-built genre among vibe coders. The blueprint made allergy safety its own engine, with a fail-closed rule for missing ingredient data, so the AI is never the last line of defence.
“…when a client disputes scope there’s no record of what they signed off.”
The “I tried to build this in Notion” classic. The blueprint added an activity log that can’t be edited after the fact: the freelancer’s evidence in a scope dispute.
Each run lists its runtime and model profile, and also ships a dev-brief for human teams. New runs are added over time.
The blueprint you approve is the blueprint your builder receives — no mid-project spec drift, no lock-in, no metered credits. MIT-licensed, zero npm dependencies, everything is Markdown. It runs on the AI coding subscription you already have.