product-init

AI sets the goal. product-init makes sure you’re shooting at the right one.

“Vibe-coding without a product spec isn’t moving fast. It’s building the wrong thing at the speed of AI.”

Codex ships /goal. Cursor ships /build. Every AI tool now moves faster.
43% of startups still die from the same cause: wrong product.

product-init is the gate before the gate — 9 hard stops between your idea and your deploy, each blocked by a Python audit script with real pass/fail criteria. CRITICAL findings stop the pipeline. No --skip flag.

Works on Claude Code, Codex CLI, and OpenClaw/Hermes.

How it works

/product-init "build an HR assessment tool"

Three questions. AI drafts the rest. 9 gates run in sequence.

Gate 1  Discovery Constitution   JTBD + four-risk model
Gate 2  Statement of Work        Shape Up appetite + PR-FAQ
Gate 3  Design                   Every screen maps to a job from Gate 1
Gate 4  Build                    Commit-to-AC, no orphan TODOs
Gate 5  QA                       Unit + integration + E2E — all green
Gate 6  UAT                      Real human signs off on real URL
Gate 7  Deploy                   HTTP 200 to prod, smoke job, rollback drill
Gate 8  Handoff                  ADRs + runbook + DEBT.md — a contract, not a README
Gate 9  Warranty                 72h monitoring window: error rate, latency, uptime

Gate 1 is the one that matters most. It asks: who gets fired if this fails, what job are they hiring it for, and what does failure look like in production?
That’s the goal you’re shooting at. Everything else is build speed.

Install

curl -sSL https://raw.githubusercontent.com/mturac/product-init/main/install.sh | bash

That’s it. The script detects which AI tools you have installed (Claude Code, Codex CLI, OpenClaw) and installs there automatically. Works on all of them at once if you have multiple.

Usage

# Bootstrap a new project
python3 scripts/orchestrator.py --project-dir /path/to/project init "your idea"

# Run a specific gate
python3 scripts/orchestrator.py --project-dir /path/to/project gate 1

# Run all audits
python3 scripts/orchestrator.py --project-dir /path/to/project audit --json

Every audit accepts --project-dir and --json. JSON output: { findings: [...], exit_code: 0|1 }.

Runtime support

Runtime	Adapter	Install dir
Claude Code	`runtime/claude-code.md`	`~/.claude/skills/product-init`
Codex CLI	`runtime/codex.md`	`~/.codex/skills/product-init`
OpenClaw + Hermes	`runtime/openclaw.md`	`~/.openclaw/skills/product-init`

Orchestrator auto-detects path: $PRODUCT_INIT_SKILL_DIR → ~/.codex/ → ~/.openclaw/ → ~/.claude/.

What ships at the end

PRODUCT.md — golden path, persona, outcome metric, kill criteria
SPEC.md — scope, acceptance criteria
PLAN.md — Shape Up pitch, appetite, deferred list
TASKS.md — golden path tasks only (filter blocks scope creep)
COMPETITIVE_BENCHMARK.md — v0/Bolt/Lovable/Railway targets
DEBT.md — every TODO/FIXME named and owned
UAT_REPORT.md — signed off, sha256-tagged
HANDOFF.md — ADRs, runbook, rollback, credentials vault link
.github/workflows/ci.yml — audit jobs as required checks

Demo

HR assessment tool built in one session with product-init:

Editorial landing page: “Hire on evidence, not on a feeling.”
Dark cinematic interview room — live AI sessions
Dashboard with scored candidates
PDF reports across 4 dimensions

Live: https://demorpoject.vercel.app

CI

.github/workflows/dogfood.yml — self-audits on every push:

Gate 1 + Gate 2 must exit 0 on known-good fixture
Orchestrator init, audit, gate subcommands tested
Runtime adapters frontmatter validated
install.sh executable check
Python syntax check on all 23 scripts
$PRODUCT_INIT_SKILL_DIR env var override verified

Research basis

Source	Applied at
CB Insights 2024 (43% PMF failure)	Gate 1 hard block
Christensen JTBD	Gate 1 Q2
Cagan four-risk model	Gate 1 Q7
Basecamp Shape Up	Gate 2 appetite + scope
Amazon PR-FAQ	Gate 2 user narrative
Torres Continuous Discovery	Gate 1 Q13
Ries Lean Startup	Kill criteria + BML loop

Full citations: references/research-evidence.md