Start

The task: $ARGUMENTS

PY is python on Windows and python3 elsewhere; use py -3 only if both fail.

1. Restate first

Your first output is this block, with zero tool calls before it. Take Phase and NEXT from the [yah] SessionStart block already in context.

Task   <the task in one line>
Phase  <label and title, or "no plan">   NEXT <the NEXT line, or none>
First  <at most one targeted search, then the edit>

No [yah] block: write Phase unknown.

The [yah] block and any task text already in context are enough: never re-read STATE.md or run bd show or which bd for them. If bd is truly needed, run store.bd from where.py --json, quoted (null: skip bd). Do not read docs/, documents/, whole plans or files over 200 lines to orient.

2. Recall

With a plan phase (or unknown), run PY "${CLAUDE_SKILL_DIR}/../../scripts/where.py" --json --no-gh. Read the phase's log, then only the ## Decisions and ### <label> sections of state.plan.spec (grep -n "^#" finds them; none: say so in one line).

Query with the task's key nouns, 12 words or fewer, no double quotes:

PY "${CLAUDE_SKILL_DIR}/../../scripts/brain.py" recall --json "<key nouns>"

• JSON printed: matches are the notes that fit, best first. Add Notes <slug>: <why it matters here> per note that applies. Open <brain_dir>/<slug>.md only when its TL;DR is not enough.
• Nothing printed: add this line once per session: "This repo has no brain folder for durable project facts. Create one (docs/brain unless config.json sets brain_dir)?" Run brain.py init only if the user says yes.

3. Begin

Do the First step without waiting.

A bug fix adds a negative regression guard in every test suite that covers the changed code: assert the bad output cannot appear (e.g. assertNotIn("**", out)), not only that the good output does.

If the task clearly takes several steps or sessions and there is no plan, suggest /yah:phases once, in one line.