Status: Early. Used by the author on real projects. Public for transparency, not actively soliciting users. Feedback welcome via Discussions; no support promises.
Product development with direction.
A vendor-neutral product development framework. Compass holds the shape of work from problem → ship → measure → learn, with AI tools playing roles across the lifecycle.
A markdown-based framework that any AI tool can read. The framework lives in compass/. Tool-specific wrappers (Claude Code, OpenAI Codex CLI, others) are thin and point at the same source of truth.
- Every initiative is a bet. Foundation product, OKRs, features, architectural initiatives — all measurable bets with a hypothesis, key metric, and an outcome: won / learning / inconclusive.
- Bets contain stories. Stories contain implementation, tests, fixes, ops.
- Roles, not job titles. 13 product roles played by AI tools loading the right context at each phase.
- Agent-agnostic by design; defaults built in. Pick any agent per role (
claude,codex,openai,gemini,deepseek,codestral,custom) from the registry incompass/config.yaml. Default = Claude implements, Codex reviews — independent models catch what one would miss (empirically validated). Per[agent-agnostic-role-assignment](v0.3.8). - Discipline holds always. Full review on every PR, no shortcuts under pressure.
- Decisions, Risks, Issues logged at every stage (DRI logs).
- Compass scans your product like Snyk scans your code. A continuous quality scanner runs across six SDLC phases — Product, Architecture, Build, Production Ready, GTM, Operate — and produces findings, not failures. Each finding has severity (Critical / High / Medium / Low) + confidence + location + reason + fix. Measurement is automatic (no manual self-assessment). Suppressions are explicit, justified, logged in DRI. Owners decide; the scanner informs.
17 workflows, grouped by when you reach for them. Several Observe workflows are auto-invoked by others — marked [auto] and rarely called by hand.
Sequenced. Run in order on a new repo.
/setup-product → Foundation product bet (PM + Researcher)
/setup-foundation-architecture → Foundation architecture bet + data model (Enterprise Architect)
/create-bet-portfolio → MVP wedge: 3-6 stub briefs + dependency graph (PM + Researcher)
Define what a bet is, design it, decompose into shippable slices.
/create-brief → New bet — fresh OR promote portfolio stub (PM + Researcher)
/create-bet-architecture → Bet-level technical strategy (Architect + Enterprise Arch)
/create-story → One shippable slice under the bet (PM, +Designer/UX Writer if UI)
Do the work. Build for stories; the others for the reactive cases.
/build <story> → Engineer implements + Codex reviews + Architect compliance
/fix <ticket> → Bug flow (Support → Engineer → Codex)
/triage <alert> → Incident response (Engineer + Support + PO awareness)
/ops <change> → Infra / config / non-code changes (Enterprise Arch + Codex)
You invoke /status, /scan, /metrics on demand. /plan, /dashboard, /measure typically run themselves.
/status → Project Manager's rolling status
/scan <bet> → Snyk-style continuous quality scanner — 6 SDLC phases
/metrics → Outcomes (won/learning/inconclusive) + open-findings posture
/plan → Living time-bound schedule (run manually or via cron)
/dashboard [auto] → Single-file HTML view of all living artifacts
(refreshed by /scan, /metrics, /plan, /status)
/measure <bet> [cron] → Cron-driven bet outcome resolution
Phase transitions: flip the artifact's
status:field directly (proposed→approved→in-build→shipped→ etc.). No canonical "advance" command — that's what status fields are for.
If you reuse a folder path for a new Compass project (delete + recreate at the same path), AI tools may carry stale context from the prior project. See SETUP.md → "Starting fresh at the same folder path" for the cleanup steps.
Read SETUP.md.