cr-production-pipeline

name

cr-production-pipeline

description

Run the canonical Campaign Receipts production pipeline end-to-end for BOTH programs on the @CampaignReceiptsYoutube channel — SEALED 2016 (145-promise audit) and CR new-news (single-race finance receipts). Spawns one specialist subagent per stage; each specialist's persona IS the gate. No council REVISE loops, no story-score-100 binding gates. Use when the founder says "produce SEALED promise N" / "make a CR new-news piece on RACE" / "make me a short on TOPIC" / "run the pipeline" or after a topic brief lands. Reads eng/PIPELINE-STEPS-AND-OWNERS.md as the canonical step→owner table and .claude/skills/cr-production-pipeline/stages.yaml as the machine-readable stage list.

Campaign Receipts Production Pipeline Skill

This skill orchestrates the canonical 31-stage pipeline documented in companies/campaign-receipts/eng/CR-PRODUCTION-PIPELINE-v4.md. It spawns persona subagents stage by stage, enforces hard-fail gates between them, and stops at the points where the founder must intervene (manual YouTube Studio upload, council overrides, KILL verdicts).

Single skill, two programs: SEALED 2016 (parchment audit) and CR new-news (navy single-race receipts). Selected via --program sealed | cr-new-news. Same gates, same QC, same /watch passes, same pipeline-state.py driver — only cosmetic + playlist + thumb template differ.

---

When to use

• Founder says "produce SEALED promise #N" / "make a SEALED LF on TOPIC" / "make a SEALED short on TOPIC"
• Founder says "make a CR new-news piece on RACE" / "build the Massie/Bush follow-up" / "do the next race"
• A topic brief lands in briefs/<date>-<slug>-topic-brief.md
• After analytics review on a prior piece says "do the next one"
• Any shorts work — re-cut, audit, batch-produce, single short. Direct invocation of cut-shorts-v2.py / produce-short-generic.mjs / render-text-short.mjs is forbidden.
• Any re-render / re-splice — surgical fixes go through the skill so Stage 19 / 24 / 25 / 26 gates run.
• Any metadata push to YouTube (youtube-upload.py --update-meta) — must enter via Stage 30 so the post-publish /watch QC runs.

NO DIRECT SCRIPT INVOCATION

Founder lock 2026-05-24 (after an agent shipped 3 broken shorts by calling cut-shorts-v2.py directly): the skill is the only entry point for any stage that has a downstream hard-fail gate. Even surgical "I just want to fix one frame" work goes through the skill — pipeline-state.py resume --from-stage N exists for this.

---

What the founder needs to supply

Minimum input to start a piece:

1. Topic brief — what's this piece about? 2-5 sentences. The story-investigator (Stage 2) extracts tension from this.
2. Slug — kebab-case:
   • SEALED: sealed-NNN-<topic> (e.g. sealed-018-china-currency)
   • CR new-news: cr-<race>-<politician>-<seat> (e.g. cr-massie-gallrein-primary)
3. Program — sealed or cr-new-news.
4. Mode — longform (with Betsy/Jessica VO) or short (text + music + b-roll, NO VO per 2026-05-24 lock).
5. (Optional) Receipt terms — <!-- RECEIPTS: china, treasury-report, fec-id --> header for the fact-check gate.

If any of #1–4 is missing, ASK before starting any stages. Do NOT guess.

---

How the skill runs (operator manual)

1. Initialize state

python3 scripts/pipeline/pipeline-state.py init \
  --slug <slug> --program <sealed|cr-new-news> --mode <longform|short> \
  --brief briefs/<date>-<slug>-topic-brief.md

Creates _review/pipeline-state-<slug>.json with all 31 stages as PENDING. LF-only stages auto-SKIPPED for short mode via only_when.

2. Loop until done

NEXT=$(python3 scripts/pipeline/pipeline-state.py next --slug <slug>)
echo "$NEXT" | jq .
# stage_id, name, kind, persona, job, inputs, output, tool, model, cost_cap_usd,
# blocking, override_flag, never_overridable, pause_reason, halt_on_verdict

For each stage:

• kind: persona → spawn Agent subagent loaded with the .md at persona as system prompt. Pass job + inputs. Verify it writes to output.
• kind: tool → run the Bash command in tool. Verify exit 0 and output exists.
• kind: gate → run tool (optionally with persona for human verdicts). On non-zero exit:
  • never_overridable: true → STOP, surface violation, escalate to founder.
  • override_flag set → ASK founder for explicit override; log to council report.
• kind: manual → pause and surface pause_reason to founder. Wait for action.

Then mark:

python3 scripts/pipeline/pipeline-state.py mark --slug <slug> --stage <id> --verdict PASS

3. Status / resume / re-roll

python3 scripts/pipeline/pipeline-state.py status --slug <slug>
python3 scripts/pipeline/pipeline-state.py resume --slug <slug> --from-stage 24

4. Cache-hit detection

next auto-marks stages PASS if their declared output path already exists on disk. The loop is idempotent — re-running after a kill picks up where it left off.

---

Persona subagent spawning

Founder rule 2026-05-24 (NEVER OVERRIDE): the main orchestrator must spawn an Agent subagent loaded with the persona's .md for every persona-named stage (script, council, QC /watch, storyboard, photo selection, generative prompts). The orchestrator does NOT do persona work itself.

Founder rule 2026-05-25 (NEVER PRE-FILTER): the orchestrator must NOT pre-filter the design space for a specialist. No "pick from these 4 options" menus to founder when a specialist owns the call. Spawn the subagent with the LIVE artifact (title, script, storyboard) + open brief + ALL options visible. The specialist decides. Founder sees the result + can re-brief if wrong.

Founder rule 2026-05-25 (PUSHBACK = RE-BRIEF): when founder pushes back on a specialist's output ("shouldn't it be X?"), that is NOT a question the orchestrator answers. Re-spawn the subagent with the founder's pushback as new context + open decision space + instruction to DECIDE the right answer (which may or may not be X). Trust the specialist's judgment over the orchestrator's tactical instinct.

Specific patterns:

• Thumbnail designer always reads the LIVE TITLE first; composition is theirs to pick (solo / 2-face collision / VS / 3-way split / geographic / faceless). The orchestrator does NOT pre-pick portrait OR layout.
• Title strategist spawn with episode facts + keyword space + brand pattern. Orchestrator does NOT suggest titles.
• Score composer spawn with locked VO + storyboard. Orchestrator does NOT suggest where ta-da hits or which instruments.
• Storyboard director spawn with locked script. Orchestrator does NOT pre-pick which shots get Wikipedia vs Remotion vs Kling.

See portfolio memory feedback_orchestrator_never_overrides_subagent_specialty.md for the full doctrine + the Rabb PA-3 incident that triggered this lock.

Spawn pattern (kind=persona):

Agent({
  subagent_type: "general-purpose",
  description: "<stage name>",
  prompt: """
You are the <persona name>. Your full system prompt is in <persona path>.
Read that file, then perform your job for slug <slug> (program=<program>, mode=<mode>).

JOB: <job from stages.yaml>
INPUTS: <inputs from stages.yaml — read these files before doing anything>
OUTPUT: write your artifact to <output from stages.yaml>

Do not skip steps. Do not produce artifacts outside your declared OUTPUT.
"""
})

For QC /watch stages (25, 26), use a qc-engineer subagent that loads personas/qc-engineer.md AND invokes the /watch skill on the master.

---

Per-program deltas

Step	SEALED	CR new-news
Slug pattern	sealed-NNN-<topic>	cr-<race>-<politician>-<seat>
Thumb template	sealed-short / sealed	cr-new-news-short / cr-new-news
CTA card	145 PROMISES · SEALED2016.COM	campaignreceipts.com/race/<slug>
Playlist	Sealed2016 (auto-resolved)	CR new-news
Story-investigator framing	"What did this promise cost or deliver?"	"Who paid for this race? What did they buy?"
All gates / personas / QC	Identical	Identical

The orchestrator passes --program through to stages that branch on it (mainly Stage 20 thumb + Stage 30 metadata).

---

Doctrine (founder lock 2026-05-25)

Pipelines are sequences of TASKS assigned to SPECIALIST SUBAGENTS. The specialist's persona IS the gate — its output IS the artifact. No council REVISE loops, no story-score-100 binding gates, no fact-check-as-blocker. If the founder doesn't like the result, founder calls it. See feedback_no_council_gates_assign_to_experts.md.

Hard MECHANICAL gates (catch defects no specialist can detect)

These are NEVER OVERRIDABLE because they check regex / ffprobe / OCR output, not subjective quality:

Gate	Stage	Script	Catches
TTS readability lint	13	elevenlabs-tts.py --lint-only	Bare AIPAC, banned phrases, URL leaks, sentence >30 words
Storyboard validator	19	validate-storyboard.py	Missing files, wrong-aspect, schema errors
Master visual QC ×2	25, 26	qc-visual-master.py + /watch	Wrong b-roll, garble, mismatched stamps
Shorts no-VO	27	qc-shorts-no-vo.py	Any speech in shorts audio

Advisory passes (NEVER block re-renders)

These produce reports the founder reads + decides on. They DO NOT trigger loop-back to earlier stages:

Pass	Stage	What it does
Research pack	6	Owns fact accuracy AT SOURCE; not a downstream gate
Council script review	11	Generates report; founder reads; specialist subagents at Steps 7-10 were the real gates
Pre-ship council panel	28	Generates report; qc-engineer at Stages 25+26 was the real gate

REMOVED 2026-05-25 (council-loop incidents on Rabb PA-3 episode)

• --founder-override-script-council flag → no longer needed; Step 11 is advisory
• --founder-override-master-council flag → no longer needed; Step 28 is advisory
• "story-score 100/100 binding" gate → optional self-check rubric only
• "copy-lock requires 4 upstream PASSes" gate → cache marker only
• "Step 2.95 fact-check-as-binding-gate" (briefly added, immediately removed) → moved into Step 6 research-pack subagent's job
• "Step 2.9b hook-strength gate" (briefly added, immediately removed) → moved into Step 10 MrBeast subagent's prompt

---

Pause-and-wait stages

Stage 29 (kind: manual) pauses for founder Studio upload. Surface pause_reason; wait for new video ID(s) in eng/youtube-meta/studio-uploads-pending.json.

Stage 31 (72h analytics) has a pause_reason — wait 72h after Stage 30. Safe to defer to a separate session; state machine survives.

---

Files this skill creates / requires

Reads (source of truth):

• eng/CR-PRODUCTION-PIPELINE-v4.md — prose doctrine
• .claude/skills/cr-production-pipeline/stages.yaml — machine-readable stage list
• personas/*.md + personas/council/*.md + personas/viral-panel/*.md

Writes (per-piece):

• _review/pipeline-state-<slug>.json
• briefs/<date>-<slug>-{story-extraction,validation,virality}.md
• research/<date>-<slug>-{sources,fact-check-pre}.md
• content/scripts/<slug>-v{1..5}-*.md
• eng/storyboards/<slug>{,-prompts,-broll-selections}.json
• _build/<slug>/{vo-v2.mp3, music_bed.m4a, thumbnail.jpg, clips/, _rem-*.mp4}
• public/longform/<slug>.mp4 OR public/shorts/<slug>.mp4
• _review/qc-visual-<slug>-pass{1,2}.md
• reports/council/<slug>-{<date>,master}.md
• reports/episodes/<slug>-72hr.md

Touches (shared aggregates):

• eng/PUBLISHED-YOUTUBE.md
• scripts/.external-costs.jsonl
• eng/youtube-meta/studio-uploads-pending.json

---

Sister skills

• banana-pro-director + cinema-worldbuilder — authoritative for visual prompt craft
• /watch — invoked by qc-engineer at Stages 25, 26
• /remember — capture new failure modes (becomes a row in v4.md "Failure modes locked into doctrine")

---

When NOT to use

• One-off Q about a published video → just answer
• Typo in PUBLISHED-YOUTUBE.md → just edit
• Analytics on a live video → use youtube-analytics-snapshot.py directly
• Doctrine / persona file edits → just edit the .md

The pipeline is for PRODUCING + SHIPPING content. Maintenance edits go around it.