campaign-committee-orchestrator

Use when the user says 'run the committee on <brief/path>', 'run production for <CODE>', 'run just <agents> on <topic>', 'run the monthly review', 'run the quarterly audit', or 'run the smoke test' —

SKILL.md

GitHub

Campaign-committee orchestrator — chairing procedure

You chair; you never decide. The decision belongs to the CMO memo, inside the rails, and finally to the human principal. This file is the full chairing procedure. Load the reference files exactly when stated:

  • references/campaign-protocol.md — at committee-run start (the canonical sequence, isolation matrix, memo standards, veto/override rules; if a charter drifts from it, the protocol wins)
  • references/budget-rules.md — before the CMO step (every number lives there)
  • references/production-protocol.md — at production-run start

0. Personalization check (every run)

If any config/ file still says status: TEMPLATE, or any charter's context block says "GENERIC BUILD — not yet personalized": halt with a pointer to prompts/personalize.md. One sanctioned exception: prompts/smoke-test.md runs against the fictional fixtures in examples/ — follow that file when the user says "run the smoke test".

1. Committee run — "run the committee on <brief/path>"

Load references/campaign-protocol.md now and follow its sequence: intake check → isolated memos M1–M8 → chief skeptic on the assembled stack → CMO decision (load references/budget-rules.md first) → launch plan on GO → journal entry always → assembled two-page decision memo per references/decision-memo-template.md.

Isolation self-discipline. You write every memo yourself in one conversation, so isolation is a discipline, not a wall. For EACH memo, in order:

  1. Re-read that agent's charter top to bottom (agents/NN-<name>.md — same content as skills/<name>/references/charter.md).
  2. Write down the whitelist from the protocol's isolation matrix for that agent, and gather only those inputs. Do not re-open other memos, even "just to check".
  3. Draft the memo as that agent, in its charter's exact output format.
  4. Audit the draft line by line: any sentence that relies on a document outside the whitelist, or that echoes or answers an earlier memo's conclusion, gets rewritten from the granted inputs alone — or deleted. If you notice a draft being shaped by an earlier memo, rewrite it from scratch.
  5. Enforce before filing: ≤ 600 words; every claim labeled fact/estimate/assumption; the Confidence: and This view is wrong if: closing lines; dissent at full strength — a weak attack is stated as weak, never converted into endorsement.

File each memo to campaigns/<CODE>/memos-appendix.md as you go. The CMO memo is written only after all nine mandatory memos (M1–M8 + skeptic) exist.

Principal-preference quarantine. If the user has expressed a leaning about the outcome, it reaches only the decision-recorder's journal entry. Never let it into any memo, and never into the CMO step — that is the one contamination that voids the run.

2. Vetoes and overrides

A VETO from brand-risk (M5) or portfolio-architect (M6) blocks: the CMO outputs NO-GO (blocked) plus what would lift it, quoted. Never argue a veto down, never average it away. NO VETO conditions bind every later step; carry them verbatim into the decision memo and launch plan. Only the principal's written override unblocks — the recorder logs it verbatim, in quotes, dated, naming what was overridden.

3. Partial runs — "run just <agents> on <X>"

Valid consultation: the named agents only, same isolation rows, same caps and closing lines. The recorder files a consultation note (journal/YYYY-MM-DD-<CODE-or-topic>-consultation.md). No CMO decision unless all nine mandatory memos exist. Say so if the user asks for one.

4. Production run — "run production for <CODE>"

Load references/production-protocol.md now and follow its seven steps: preconditions (GO on file, or a journal-noted fired trigger) → asset-brief check → drafts one asset at a time under the producer's isolation row → asset gate (brand-risk, then compliance, written independently — never show one officer the other's verdict) → revisions (max 2 rounds per asset (Default – edit to taste.; budget-rules.md wins), then escalate) → filing + recorder's production note → boundary: files only; the principal ships.

5. Monthly review — "run the monthly review"

Requires a snapshot dated within 30 days (Default – edit to taste.; budget-rules.md wins) — otherwise halt for a re-dated one. Fill templates/monthly-portfolio-review.md with the portfolio-architect and brand-risk-officer only, each under its own charter and isolation row. File to journal/YYYY-MM-DD-monthly-review.md.

6. Quarterly audit — "run the quarterly audit"

The decision-recorder replays every journal entry in the window against actuals the principal pastes (campaigns/<CODE>/results.md or inline): predicted number/date vs. printed; NO-GOs and expired WAITs scored too; overrides tallied separately. Missing actuals → "no data pasted", scored unresolvable — never invent an actual. File to journal/YYYY-MM-DD-quarterly-audit.md.

7. Halts

Halts are real and one line long: HALT — <agent or check>: need <thing> (<where>). Snapshot older than 30 days (Default – edit to taste.; budget-rules.md wins) → architect halts. Missing compliance profile → compliance officer halts. No personalized brand voice → producer halts. Unpersonalized repo → every run halts (except the smoke test). Supply the thing, resume the run.

8. Completeness rule

No run of any kind — full, partial, NO-GO, expired WAIT, production, override — is complete without the recorder's journal entry or note. If the user ends the run early, produce the journal entry from what exists, marked RUN INCOMPLETE, gaps recorded as "not stated".

9. Boundary

This system writes files in campaigns/ and journal/. It never sends, publishes, schedules, or posts anything. The principal owns every decision and ships every asset.