Ce Produce

Content pipeline

1. Input: an APPROVED brief item (require_approved - drafting counts as work the user pays attention for, so briefs are gated before drafting). A user may also hand a manual brief. Register it before anything else: confirm scope in-session, then create_item(kind="content-brief", ...) with the brief text as the body - items are born proposed, never any other status - and record the in-session approval via the contract CLI so require_approved passes. Never save a hand-written file into briefs/ and never set a status at creation: a brief born drafted with no approvals can neither be approved nor published, and jams the pipeline until python3 -m core reset-to-proposed repairs it. A brief this pipeline will draft still starts proposed and reaches drafted only through its approved lineage (step 4). 1b. Decision check, BEFORE registering a manual brief with create_item. Search the brain's decision memory for the topic: PYTHONPATH="$CLAUDE_PLUGIN_ROOT/lib" python3 -c "..." snippet importing core.decisions - search(<brain>, [<topic>, <target keyword>]). It returns prior decisions newest first with date, choice, scope, item, and the rationale. A hit whose choice is rejected means a human already refused this piece. Exactly two paths are allowed, never a third:
   • SKIP it: tell the user the topic was rejected on because , and stop before creating anything.
   • Or CREATE it with a line in the brief body reading "previously rejected on because ; proposing again because ". What changed must be concrete - a new signal, a search- demand shift, a product change - never "worth another look". Silently re-registering a rejected topic is forbidden. A hit with any other choice is context for the brief, not a blocker. An already approved brief arriving from the queue has cleared this check at proposal time; do not re-run it there.
2. Stages, sequential agents (each receives profile path + prior artifacts): ce-researcher -> ce-writer -> ce-brand-auditor -> ce-seo-aeo -> ce-qa -> ce-editor.
3. Save the final draft as .draft.md next to the brief (frontmatter: title, slug, meta_description, focus_keyword, schema block, and capsule: verified - written only after ce-qa's capsule hard check passed AND ce-editor's Capsule confirmation line is present; this is the handoff note onsite-publish requires before publishing). Save the research pack + QA log under runs/YYYYMMDD-produce-/.
4. PYTHONPATH="$CLAUDE_PLUGIN_ROOT/lib" python3 -m core status <brief-path> drafted --actor agent. Then invoke ce-image for the featured-image step. Publishing is onsite-publish's job (separately gated).
5. Rules from the profile override everything: voice, banned phrases, rulebook. House defaults if profile is silent: answer capsule up top, no em-dashes, active voice, statistics cited inline with live links.
6. Never edit brain frontmatter directly. The contract CLI is the only write path for status and approvals.
7. At each stage boundary, append a one-line progress marker with a UTC timestamp to the run report file before starting the stage - headless runs are watched by tailing that file, not a terminal.

Comparison briefs (brief_type: comparison)

When the brief's frontmatter carries brief_type: comparison (the field is additive; absence means explainer - see docs/site-repo-contract.md), the same six stages run with comparison-specific guidance layered on:

• ce-researcher: verify EVERY claim about a third-party product against that product's live public pages - pricing and features change, so each claim destined for the comparison table needs a fetch-verified source URL and a checked-on date in the research pack. Never from memory.
• ce-writer: produce the X-vs-Y structure - the answer capsule states the honest one-line difference between the two; the comparison table carries only rows the research pack verified; a "who should pick which" section treats the competitors fairly, naming where each is the better pick.
• ce-qa: two comparison-specific hard checks on top of the standard ones: (1) no unverifiable competitor claims - a row whose cells cannot be matched to a fetch-verified source is dropped, never guessed; (2) no disparagement - factual differences only, no adjectives about competitors.

Rationale: X-vs-Y and listicle shapes are the most-cited content shapes in AI search (https://www.position.digital/blog/digital-pr-tactics/).