Pm Context System

PM Context System

Memory built one approved line at a time. The moment something worth keeping happens, one proposal — never a questionnaire, never a silent write.

Verification gates (defined first; output is blocked until all pass)

• G1 — Single-line proposal form: every capture proposal is exactly one line: Log to memory: "<fact/decision + why>" — yes/edit/skip. Multi-line proposals, batched questions, or any interview fail the gate.
• G2 — Consent before write: nothing is written to any file without a yes or edit on its specific proposal. Silent writes fail; a skip is final — that item is never re-proposed.
• G3 — Event-driven, deduplicated: proposals fire on actual session events (a decision lands, a stakeholder fact surfaces, state changes) and never re-propose what's already logged. The end-of-session sweep proposes only the unlogged remainder.

Steps

1. Watch. During any session, when a decision is made, a stakeholder fact surfaces, or project state changes, emit the one-line proposal tied to that specific event — then continue the task. Nothing noteworthy → total silence; an empty check contributes zero lines.
2. Scaffold on first yes, silently. If context/ doesn't exist when a proposal is approved, create it without announcement: INDEX.md (read order + one-line file summaries), STAKEHOLDERS.md, DECISIONS.md (append-only), STATE.md. Write the approved entry where it belongs; add one pointer line to CLAUDE.md: "At session start, read context/INDEX.md first."
3. Read order at session start: INDEX → STATE → the other two only when the task touches a person or reopens a decision. Progressive disclosure; never bulk-load all four.
4. Sweep on request or session end. "Update context" → one pass over the session for unlogged decisions/state changes; each gets its own one-line proposal. Already-logged items are skipped — the sweep is a diff, not a replay.
5. Maintain when touching files: merge duplicate entries (keep newer framing); STATE items idle 30+ days → propose archiving (never silently delete); contradicted entries → flag [STALE?] and ask; DECISIONS.md corrections are new entries referencing the old, never edits.

Output format

(mid-session, a decision just landed)
Log to memory: "API pricing: usage-based over flat — flat penalizes low-volume pilots" — yes/edit/skip

(after "yes", first time only — silent scaffold, then:)
Logged to context/DECISIONS.md.

(session end, "update context")
Log to memory: "STATE: billing migration blocked on legal review of usage terms" — yes/edit/skip
(nothing else — the pricing decision above is already logged)

Hard rules

1. Never interview. Five questions up front is the failure mode this skill replaces — one proposal per event, tied to what just happened, always skippable.
2. Never write without consent, and never re-propose a skip. Trust in the memory system is the product; one silent write ends it.
3. Never fabricate memory: entries record what was actually said this session — no inferred stakeholder opinions, no filled-in rationale the user didn't give (an edit exists for exactly that).
4. DECISIONS.md is append-only. History that can be rewritten isn't history.

Limitations

• Memory lives in repo files; it persists per project and syncs only via git — nothing crosses machines or projects on its own.
• A session must read the files for memory to work; the CLAUDE.md pointer makes that reliable in Claude Code, best-effort elsewhere.
• Quality depends on the ritual actually running — skipped sessions create silent gaps, not errors.
• STAKEHOLDERS.md holds professional context only: no health, private-life, or anything a stakeholder wouldn't expect written down — such proposals are never made.