▶ claude -p (cwd: /home/app/projects/content-engine)
You are running the Living Neurodivergent content AUTOPILOT for this idea:

> "create a reel about the relationship with adhd"

Execute every step in order. Save outputs to the standard paths. Do not ask me anything — make decisions yourself based on the brand's data files. This is a single fire-and-forget run that ends with a publish-ready pack.

STEP 0 — Slug only (FORMAT IS LOCKED TO: reel)
- The user has chosen the format: **reel**. Do NOT override it. Treat this as a hard constraint, not a suggestion.
- Still read CLAUDE.md, performance/hooks-*.json, performance/*-digest.md, and products/PRODUCTS.md to inform the strategist and writer downstream.
- Pick a SLUG (lowercase, hyphenated, 2-5 words from the topic).
- Use today's date for {today}.

STEP 1 — Strategist (write the brief)
- Read .claude/skills/strategy-protocol/SKILL.md.
- Produce briefs/ad-hoc/{today}-{slug}.md following that skill's brief template.

STEP 2 — Writer (produce the draft)
- Read .claude/skills/writing-protocol/SKILL.md.
- Read products/PRODUCTS.md again so the CTA references real modules.
- Write drafts/{today}/{slug}.md per the writer protocol. Apply per-format offer-tied rules. NO inline citations. Pain-driven CTA. Sentence rhythm.

STEP 3 — Editor (verdict)
- Read .claude/skills/editing-protocol/SKILL.md.
- Write approved/{today}/{slug}.md with the verdict APPROVED, NEEDS_FACT_CHECK, or NEEDS_REWRITE.
- If NEEDS_REWRITE: silently go back to STEP 2 (one rewrite only, then proceed regardless).

STEP 4 — Fact-checker (only if editor returned NEEDS_FACT_CHECK)
- Read .claude/skills/fact-checking-protocol/SKILL.md.
- Write fact-checks/{today}/{slug}.md.
- Then re-run the writer to integrate hedging language for any PARTIALLY_SUPPORTED / UNSUPPORTED claims.
- Then re-run the editor. Expect APPROVED this time.

STEP 5a — Visual director (production brief)
- Read .claude/skills/visual-direction-protocol/SKILL.md.
- Write visual-briefs/{today}/{slug}.md.
- For carousels and singles: ensure the brief contains a `## AI generation prompt` section with one `**Slide N:**` fenced code block per slide (1 for single, typically 8 for carousel). Step 5b reads these blocks verbatim.
- For reels: one general AI video prompt only.

STEP 5b — Generate images (carousel + single only; reels skip this step)
- Read .claude/skills/image-generation-protocol/SKILL.md.
- Execute it against visual-briefs/{today}/{slug}.md. The skill handles directory creation, per-slide higgsfield invocation, downloads, the brand-style suffix, idempotency, and the `## Generated assets` append.

STOP. Do NOT post to Instagram. Do NOT call the analyst or own-performance-analyst.

Final output: print one line — "Autopilot done: {format} pack ready at approved/{today}/{slug}.md + visual-briefs/{today}/{slug}.md + assets/{today}/{slug}/".



[self-heal] marked as lost — host process gone