| name | pm-mode |
| description | Use when shipping a product (post-PMF). Runs HARVEST → SHAPE → BUILD → SHIP cycles indefinitely. Each cycle harvests signals, proposes a release plan with kill-list (10% Delete Rule) and Devil's Advocate, runs the Ralph Loop with autonomy tiers (T1/T2/T3), audits drift every 5 features, ships, measures, recommends stabilization if quality slips. |
PM mode
You ship a living product through repeating release cycles. Each cycle has discipline — kill list, drift checks, autonomy tiers, stabilization triggers.
Difference from Founder mode: Founder mode is finite (3–5 cycles → graduate or kill). PM mode runs forever — every release ships, measures, harvests signals for the next.
The cycle
HARVEST → SHAPE → BUILD → SHIP, then loop.
HARVEST (15–30 min)
Collect raw signals from feedback, analytics, tickets, competitive observations. Auto-mine git log + open issues + recent PRs. Categorize each, assess thesis_impact (null / confirms / challenges / invalidates), refresh architecture.md incrementally.
Pause: review signal summary. Invalidations get prominent flagging.
SHAPE (30–60 min) — the hero phase
Propose a release plan:
- Thesis (1 sentence)
- Identity evolution — only if signals warrant. Trigger + rationale + risk.
- Features — each with
signal_id, requested_by, intent, autonomy (T1/T2/T3)
- Debt paydown (~15%)
- Kill list (≥10%) — the 10% Delete Rule. Mandatory.
- Contract implications (if
contract_mode != "none")
- Appetite — what's NOT included
- Stabilization check — auto-fire on rework>15% / blocked≥3 / resets>2 / quality<6
Devil's Advocate auto-runs against the plan. Surfaces 2–3 attacks inline. The output is the hero card — screenshottable monospace render with ✓/⚠/✗ markers.
Pause: human approves, modifies, or kills the cycle.
BUILD (async, hours)
Feature breakdown decomposes the plan into a wave (additive — don't re-implement what's done). Then the Ralph Loop:
SELECT next feature → TEST (3–7 acceptance tests) → BUILD (fresh context, max 3 resets) → EVALUATE
Autonomy tiers determine where humans checkpoint:
| Tier | Test | Build | Commit |
|---|
| T1 | Auto | Auto | Auto-commit |
| T2 | Auto | Auto | Checkpoint before commit |
| T3 | Checkpoint | Checkpoint | Checkpoint before commit |
Every 5 completed features → drift check. DISCUSS / PIVOT / CONTRACT_HALT pauses the cycle.
Pause is per-feature for T2/T3 and per-5-features for drift checks.
SHIP (15–30 min)
Contract validation (halt on unapproved BREAKING_CHANGE), architecture refresh, log release in releases.json with full metrics, stabilization-trigger check for next cycle.
Pause: present release report. The next /forge:harvest starts the next cycle.
State files (6)
| File | What |
|---|
IDENTITY.md | WHO/JOB/NEVER + Architecture Constraints + Evolution Log |
features.json | Cumulative feature backlog with signal_id, requested_by, autonomy, quality_score, build_resets, rework_of |
releases.json | Release history + current_cycle (signals, phase, approved_plan) + contract_mode |
progress.md | Cross-Cutting Patterns, Active Gotchas, Architecture Decisions, Recent Drift Reports, per-release feature learnings |
architecture.md | Module hierarchy, hot paths, blind spots, API surface, integration points |
contracts.md | Per-API/schema/event commitments, deprecation timelines (opt-in via contract_mode) |
Quality score per feature
30% Tests passed first build (10 if first try, –2 per reset)
20% No REWORK flag (10 clean, 0 if rework)
20% No BREAKING_CHANGE (10 clean, 5 if flagged + approved)
15% Contract compliance (10 compliant, 0 violated)
15% No TECH_DEBT flag (10 clean, 5 if flagged)
Avg quality_score < 6.0 across last 5 features → recommend stabilization.
Token efficiency
PM mode is a token-saving harness vs unguided agentic coding:
- Fresh context per feature — never carry context across boundaries
architecture.md as a 500-token map vs forcing the agent to re-explore 50KLOC- State files re-read between phases (~3–5k tokens) vs carrying full context forward
- Drift checks catch rework before re-implementation — the most expensive class of wasted tokens
- All agents default to Opus 4.7 (1M context). Users can downgrade specific high-volume agents (builder, test-writer, market-scout) to Sonnet by editing their frontmatter.
Target ceiling: ~60k tokens per feature build. Unguided agentic coding routinely runs 100k+.
When to switch to Founder mode
If you start receiving signals with thesis_impact: "invalidates" and the drift detector starts recommending PIVOT — your IDENTITY.md may be wrong. Drop into Founder mode for one cycle, validate the new thesis, then return.
