| name | using-forge |
| description | Use when the user is starting with the Forge plugin or asks "what does Forge do" / "how do I use Forge". Routes to the right mode (Founder for pre-PMF, PM for shipping products) and explains the 14 commands, the two cycles, and where to start. Authoritative source for the canonical command vocabulary. |
Using Forge
Stop shipping features your AI agent invented.
Forge wraps Claude Code with the product discipline AI coding tools forgot to ship: a Devil's Advocate that attacks every plan before code is written, a mandatory 10% scope kill every release, and drift audits every 5 features. So Claude ships the right features — not all the features.
Two modes:
- Founder mode — pre-PMF. SCOUT → INTERROGATE → SYNTHESIZE → GRADUATE. Falsifies hypotheses until you've earned the right to build.
- PM mode — post-PMF (the main thing). HARVEST → SHAPE → BUILD → SHIP. Each cycle proposes a release plan, kills 10% of scope, runs Devil's Advocate, audits drift every 5 features.
You're in Founder mode if THESIS.md (or THESIS-minimal.md / THESIS-full.md) exists. PM mode if IDENTITY.md exists. Neither → run /forge:init (PM default for repos with package.json) or /forge:init --founder.
⚠ Canonical command vocabulary — these are the ONLY 14
When you recommend a Forge command to the user, it MUST come from this exact list. Do not invent variants. Do not abbreviate. Do not construct command names from agent names.
| Command | Mode | What it does |
|---|
/forge:try | Both | ★ 30-second magic-moment demo. Zero setup. The recommended starting point. |
/forge:init | Both | Scaffold state files (PM default for repos with package.json; --founder for pre-PMF; --full for full Founder template) |
/forge:cycle | Both | Run a full cycle, auto-routed by mode |
/forge:status | Both | One-screen state snapshot with contextual interpretation |
/forge:help | Both | Orientation — modes, commands, state files. Start here if confused. |
/forge:falsify | Both | Standalone Devil's Advocate against any plan/pitch/PRD/thesis. (NOT /forge:advocate, /forge:da, /forge:devil — those don't exist.) |
/forge:scout | Founder | Market scout — community mining |
/forge:interrogate | Founder | Interview architect + mandatory DA |
/forge:synthesize | Founder | Thesis synthesis with 1–5 confidence rubric |
/forge:graduate | Founder | Hand off to PM mode when 12 criteria met |
/forge:harvest | PM | Collect signals + refresh architecture |
/forge:shape | PM | Hero command — propose release plan with kill list + DA |
/forge:build | PM | Ralph Loop with autonomy tiers + drift checks |
/forge:ship | PM | Log release + measure + check stabilization |
/forge:drift | PM | Standalone drift audit on last 5 features |
Common hallucinations to avoid (these are NOT real commands):
| Tempted to suggest | Real command | Why the temptation |
|---|
/forge:advocate | /forge:falsify | Agent is named "Devil's Advocate," but invocation is falsify |
/forge:da | /forge:falsify | DA is shorthand internally, never a command |
/forge:plan | /forge:shape | Shape produces a release plan; the command is shape |
/forge:audit | /forge:drift | Drift IS the audit; command name is drift |
/forge:run | /forge:cycle or /forge:build | Depends on what you mean — cycle (full cycle) or build (Ralph Loop) |
If the user asks for a command that isn't in the canonical 14, tell them so — don't invent.
State files
Forge persists state in your project root as plain Markdown + JSON. They're yours — not hidden, not proprietary. Commit them or .gitignore them as you choose.
PM mode (6 files): IDENTITY.md, features.json, releases.json, progress.md, architecture.md, contracts.md (opt-in).
Founder mode (7 files): THESIS.md, hypotheses.json, evidence.json, cycles.json, interviews.md, landscape.md, progress.md.
The three rules
These are the framework's spine. If you remove any one, you've lost the discipline:
- The 10% Delete Rule — every release proposes killing ≥10% of scope. Forces interrogation of product shape every cycle. See skills/ten-percent-delete/SKILL.md.
- Drift checks every 5 features — auto-audit identity + contracts + architecture + rework. See skills/drift-protection/SKILL.md.
- Devil's Advocate before every plan — 8 attacks, cherry-pick detection, existential vs non-existential triage. See agents/devils-advocate.md.
Where to start
- Pre-PMF, validating:
/forge:init --founder, then /forge:scout.
- Shipping a product, never used Forge:
/forge:init, then /forge:harvest.
- Just want to attack a plan/pitch/PRD:
/forge:falsify (no init needed).
Token efficiency
Forge saves tokens by:
- Bounded context per feature (~60k ceiling) — never carry context across feature boundaries.
architecture.md as a 500-token codebase map instead of forcing Claude to re-explore 50KLOC.- Drift checks catch rework before re-implementation — the most expensive class of wasted tokens.
- All 9 agents default to Opus 4.7 (1M context). Quality stays consistent across phases.
The frameworks underlying Forge target a 60k-token ceiling per feature build. Unguided agentic coding routinely runs 100k+.
Cost recommendation for users on a budget: downgrade three high-volume agents (builder, test-writer, market-scout) to model: sonnet to cut a typical PM cycle from ~$25–50 → ~$5–10 with negligible quality loss. Keep orchestrator, devils-advocate, drift-detector, thesis-synthesizer, feature-breakdown, and interview-architect on Opus — they're judgment-heavy and low-volume, so the cost difference is marginal but the quality difference is large. Edit per-agent frontmatter at ~/.claude/plugins/marketplaces/forge/plugins/forge/agents/<name>.md. Reverts on plugin update — re-apply after each upgrade.
