ac-plan — idea in, one plan file out

I/O Contract

Input	An idea, a backlog item, or a problem statement
Output	ONE plan file, _plans/YYYY-MM-DD-HHMM-<slug>.md, unstamped
Artifacts	Explorer notes in _plans/research/ — only if explorers were run
Verification	ac-polish plan <path> to fixpoint; ac-polish/references/plan-checklist.md is the bar

Doctrine: skills/ac-pipeline/SKILL.md — including the model-tier Calibration (planning runs OPUS-tier; a Calibration with a retirement measurement, not a fact to restate here). Bead and
commit canon: beads-standards and ac-pipeline/references/ BY POINTER — restated nowhere.

Before anything: the task-size floor

A trivial single-file change with no dependency structure BYPASSES ac2 entirely — no plan, no beads, no ceremony. Applying the pipeline to a nit is a defect in judgement, not diligence.

What goes in the plan — the ten-year test

Be optimistic about the model, pessimistic about what we must actually say. State only what a capable model cannot infer: our conventions, our file locations, our failure history, the specific decisions this work turns on. Every line that would be obvious in ten years is
a line to cut — the plan is graded on whether its claims are checkable, never on its length.

The citation rule. Every claim about the tree cites a path plus a quoted phrase or the finding command — never file:line, which drifts on a shared trunk before the claim does.

Procedure

1. Size the work. Below the floor → stop, do it directly. Otherwise continue.
2. Ask only the obvious questions — the ones without which the brief would be a guess; a capable read of the codebase answers the rest. Explorers are OPTIONAL, size selects them. Notes land in _plans/research/; none ran → say so.
3. Draft the vision brief in plain words a non-engineer follows — no pipeline
   jargon, and the same rule wherever the vision or plan is shown to the human —
   corrected by the human in free text, then frozen verbatim as ## Vision.
4. Seams scan. Every path in the conductor's drafted deliverable list is an object, no
   cap — ## Deliverables is written in step 5, so the scan reads the draft, never the
   plan. Derive its file list (skills/_tools/touchers.sh derive <path> plus the repo's
   test globs) and spawn one fresh reader per object IN PARALLEL, sent
   references/plan-seams-reader.md verbatim — only for a non-empty derived list (empty
   gets no reader: an empty sweep proves nothing). One row per file, silence banned,
   completeness by count; an OUT-OF-PLAN claim survives only when it names the deliverable
   it compromises, is P0–P2, and a fresh second reader confirms it (same file, § Confirming
   an OUT-OF-PLAN claim).
5. Write the ONE plan file, ## Vision first: Problem, Approach, Deliverables (artifacts, never intentions — — one-shot marks migration scratch, D4), Assumptions (what DETECTS
   the falsity, by when), Decisions (§ next step), ## Seams (after Decisions — one row per surviving finding, object by object; plan-checklist.md § 1 re-derives every object's touchers against it — an uncovered file is an unowned seam), Risk + sequence (Human gates: names every authorization gate), Out of scope, Success criterion.
6. Decisions and improvements per references/decisions.md: settled: <choice> — <why> or needs-human, each settled card quoting its ## Vision line;
   up to two opt-in improvements alongside, never folded into Deliverables on the agent's own judgement — each accepted one appends one sentence to ## Vision, quoted by its card.
7. Diet pass, bounded. Spawn a fresh orchestrator-tier reader with the plan and one question — where is the fat: the same functionality,
   experience and reliability with less or simpler. It returns cuts, each naming what replaces it, never edits. The conductor checks each claim and
   applies a cut only when nothing concrete is lost — a safeguard, a Seams finding, a Vision sentence left unmet — keeping the rest without a card; a
   question is answered by lookup; a trade it still backs rides the opt-in line (references/decisions.md § Improvements). Stop when a round applies nothing, three rounds at most.
8. Approve. Render references/approval-brief.md's brief and run ONE question round — open cards plus Approve / Change / Park. Approve runs
   skills/_tools/plan-approve.sh approve <plan> "$(git config user.name)", the ONE writer, never a hand edit. Unattended: the plan stays draft and the docket shows it waiting.
9. Stop. End with the line Next: /ac-polish plan <path> — never invoke the next stage.

The success criterion — a refusal, not a suggestion

REFUSE to emit a plan whose success criterion cannot come out FALSE. Under the heading ## Success criterion it is the Silver Bullet: one command, its expected result, run while planning, its failing output pasted in. Three ways it fails, each sends it back:

• Unfalsifiable — prose, not a command; no observation could contradict it. It is a slogan.
• Already true today — it passes now, so nothing was asserted; paste the failing output, or it is not this plan's criterion.
• Unowned — nobody is named to run the command, on what artifact, and when.

One Silver Bullet, not three. Every ## Deliverables bullet carries a "Done when:" line with concrete values: what goes in, what comes out.