decide

Decide

Turn a fork in the road into a clear, logged decision — reasoning visible, and the way you'd know you were wrong stated up front.

When to run

A real choice with stakes (a tradeoff, a hire, a vendor, a direction, a bet).
After /structure surfaces a "Decision needed."

The output (always this shape)

## Decision — <the question> · <date>
**Context** — why this is on the table now (2–3 lines).
**Options** — each with its main upside + main cost/risk (include "do nothing").
**Criteria** — what actually matters here, ranked.
**Recommendation** — the call + the why, in one paragraph.
**Falsifier** — what we'd later see that would prove this was the wrong call (Verification Gate).
**Reversibility** — one-way or two-way door? (two-way → decide fast and move).
**Owner / next step** — who does what, by when.

Process

Frame the real question. Strip to the actual choice (/structure if fuzzy). Name what is not being decided.
Lay out options honestly — including "do nothing." Each gets a real upside and a real cost. No strawmen.
Make criteria explicit and ranked — most decisions go wrong because the criteria were never named.
Recommend. Take a position with the reasoning. State the falsifier and the reversibility (two-way doors don't deserve one-way-door deliberation).
Verify the inputs (Verification Gate). Any fact or number the decision rests on → trace to source; flag unknowns as risks, don't paper over them.
Log it. On the user's confirmation, append the brief to decisions/log.md (dated) — the standing record (per "How you work with me").

Autonomy

L1–L2 — frames and recommends; the human makes the call. Logs it once the user confirms.

Guardrails (from CLAUDE.md)

Cite, don't invent — decisions rest on verified facts; unknowns are named as risks.
Log decisions to decisions/log.md.
No confidential figures or PHI/PII in the log (it's committed) — reference the source, don't transcribe.