# Session kickoff — master prompt (paste this to start any session)

> Maintained by the orchestrator per the operator's ask (2026-07-20). Update
> it when a standing steer changes; it must stay paste-ready and short.

---

# FIRST THING

Update the local repo to match the remote. In a lane worktree that means
`git fetch origin && git rebase origin/main` — never pull-merge, and never
work from the hub checkout. If you expected new remote commits and there are
none, stop and say so — do not guess. Then run the post-pull regen your
memory prescribes (prisma generate + shared build if db/shared changed)
before any tests.

Resuming a specific branch? Read ONLY that lane's opener at
`plans/kickoff-prompts/<lane>.md` — never the whole directory — and its own
state file (`plans/rethink/<lane>-handoff.md`) outranks the shared HANDOFF
for that lane.

# START SESSION

Orient from files first — this conversation carries no memory.
CLAUDE.md is already loaded for you (project instructions) — it carries the
deploy loop, the test recipe, and the precedence order when docs disagree.
Then read, in order: HANDOFF.md (top block — it carries the current ORDER)
→ plans/rethink/scope-board.md → plans/rethink/README.md (which packages are
open) → docs/DECISIONS.md (every ruling in it is in force; read the sections
that touch your task — Process & roadmap and Dead ends always).
Open only the kickoff file for the scope you are actually picking up.
Nothing else yet.

If two docs disagree, or a doc contradicts what the repo actually shows —
do NOT pick a winner and do NOT reconcile silently. Surface it: which
files, which lines, what each claims. Stale state is my call, not yours.

Then report, under 15 lines:
- POSITION — where the work stands, cited to file:line.
- OPEN — unfinished, blocked, or claimed-done-without-fresh-proof.
- CONFLICT — any doc/state disagreement (or "none").
- RISK — the landmine nobody asked about.
- NEXT — the single highest-value move + the exact oracle that proves it done.

Surface the OPERATOR-BLOCKED list from HANDOFF at the very start.

Then at most one question: the fork only I can resolve, your recommendation
attached. No real fork → say so and start NEXT.

# STANDING LAWS (these override defaults; ledgered in DECISIONS)

- **Make the agent WORK. No re-rethink.** Real jobs working beats new
  features. No design rounds unless I ask for one.
- **Simplify every screen you touch** — plain words, fewer concepts. New
  surfaces land on `--cm-*`; flag a legacy host for a sweep, never extend it.
- **Prune scar-tissue rules**, in the bot's seeds and our docs alike. Seeds
  define what the bot IS, never what it mirrors from the sender.
- **A MOCKUP IS A UI PROMISE, NOT A FEATURE LIST** (operator, 2026-07-30, said
  three times before it landed). An approved mockup binds the LAYOUT, the visual
  design and the ergonomics as strictly as it binds the behaviour. Building the
  functionality and grafting it onto the existing screen is a broken promise,
  even when every test is green and the tokens all resolve. **Never let what was
  already implemented set the shape of what ships — that is precisely what the
  redesign was asked for.** If the built screen and the mockup would not be
  mistaken for each other side by side, it is not done. Look at both. Screenshot
  the mockup and the build and compare them before claiming a screen.
- **Verify with your own oracle** — a seat's report is a claim. Deploys,
  pushes, publishes are pre-authorized; gate flips, LLM spend, Meta config
  and destructive acts are mine.
- **Spawning**: name the tier and why. Executors Opus, mechanical sweeps
  small, biggest batch per seat, state in files. Shared trees: atomic
  `git commit -- <pathspec>`, rebase with backoff, never delete index.lock,
  pushed state must build.
  **Fan out CONTROLLED output only — never creative output** (operator,
  2026-07-30, after nine of eleven parallel-authored mockups were rejected):
  *"orchestration should be done in case of a controlled output not creative."*
  A mechanical sweep, a measurement, a search, a per-file migration — fan those
  out. Design, mockups, copy, anything where consistency IS the deliverable —
  one hand, screen by screen, against a stated direction. Parallel seats produce
  work that is individually defensible and collectively incoherent, and it costs
  more. Orchestrate execution you can control; keep the creative work in hand.
- **Ledger decisions in DECISIONS**, a reversal merged into the line it
  reverses (originals in `docs/archive/DECISIONS-history.md`). Deferrals go
  on the board, never only in prose.
- **HANDOFF is a state file, not a log** — rewrite in place, delete closed
  items, re-verify every surviving line, ≤120 lines.