content-creation-framework

Bounded check runner for content artifacts. Executes activity-check scripts declared by the invoking skill (e.g. content-writer, editor, research-analyst) and returns structured JSON of findings — pass/warn/fail per check, aggregated overall status, raw stdout/stderr captured. Use when a skill needs to run its declared skill-local activity checks (citation count, structure conformance, voice match, schema validation) against a specific artifact without polluting the parent context. Per P9, runs are advisory and skill-local; the sub-agent does NOT gate progression and is not a phase or pipeline gate — it reports findings, the invoking skill and the user decide what to do. Paired with content-writer for draft-time checks and with any producer skill that declares its own check scripts.

Bounded extraction worker for source materials. Reads files and URLs, returns structured JSON of facts, quotes, patterns, voice characteristics, or credibility signals — each item attributed to its source with a confidence score. Use when another skill (e.g. research-analyst, content-validator, voice profiling) needs raw extracted evidence from a defined set of source_paths without polluting the parent context. Delegate when source count exceeds ~3 OR when source-content volume would risk filling the parent context window; for small inline-tractable extractions, the parent skill may read sources directly without delegating. Supports six modes: patterns, ingest, facts, quotes, credibility, voice-profile.

Develops the anchor documents that scope every piece of content production for a project: the `content-brief.md` (strategic intent — audience, key messages, objectives, voice direction, format, success criteria) and the `outline.md` (structural plan — section purposes, narrative arc, word allocation, fact-coverage hooks). Reads upstream research synthesis when it exists; elicits intent collaboratively when it does not. Pairs with `content-writer` (downstream consumer of both anchors) and `editor` (reviewer of drafts against the same anchors per P11). The brief and outline pass through `artifact-reviewer` for critical review before user accept. Use when the user says: "plan the article on X", "write me a content brief", "outline the iurfriend Q2 piece", "let's strategize the Trennungsjahr article", "what audience are we targeting for Y", "develop the brief and outline for Z", "I need a section plan", "scope this content", "draft an outline from the research", "set up the planning artifacts for the climate-policy article", or otherwise asks for content strategy, content brief authoring, outline authoring, or section-plan production. For drafting prose, route to `/content-writer`. For editorial review of drafts, route to `/editor`. For research gathering before planning begins, route to `/research-analyst`.

Produces a content draft (`draft.md`) against its anchor rubric (`content-brief.md` + `outline.md`) for a content project. Applies brand canonicals (`BRAND.md`, `VOICE.md`, `AUDIENCE.md`) for voice and audience targeting; integrates research facts from the project's `research/synthesis.md` and `research/facts/` with explicit source attribution (P7); follows the outline's section structure. For long outlines, shards the draft into per-section passes internally and stitches them together with smoothing — the user sees one `draft.md`, not a checkpoint zoo. Ends at `status: draft → review` only; never writes `status: greenlit` — the orchestrator performs that mutation on explicit user accept per P8 + R25 + R38. Use when the user says: "draft this article", "write the draft against the outline", "produce the draft for <article>", "start the iurfriend-q2-trennungsangst draft", "continue drafting from where we left off", "rewrite section 3 of the draft", "apply the editor's feedback to the draft", "produce a unified pass", "draft this section by section so I can review as you go", or otherwise asks to produce or revise the article body itself. For review of the produced draft, route to `/editor`. For publishing it, the orchestrator runs the publish routine (there is no content-publisher skill).

Retire a published content artifact: draft a deprecation notice next to it, propose the `status: published → deprecated` frontmatter transition (with `deprecated_at` and optional `superseded_by`), and surface the proposal for user accept (P8). Aligns with P10 (artifact lifecycle). Use when the user says: "retire this article", "mark this as outdated", "this piece is no longer relevant", "deprecate the Q1 2024 content", "we need to sunset this", "replace the old <X> article with the new one", "this is superseded by <Y>", or any signal that a published piece has reached end-of-life. Applies only to `status: published` artifacts — for `draft | review | greenlit` pieces, use the archival sweep instead (move to project-root `_archive/` with `status: archived`).

Review and edit a content draft against its anchor rubric (the content-brief.md and outline.md the work was produced under). Surfaces coherence, strategy alignment, compliance, accessibility, voice-alignment, originality, and publication-bound publish-package integrity findings; flags blockers; recommends revisions. Produces a `review-report.md` (type: content/review, status: review). Pairs with the `artifact-reviewer` skill (with the `draft_mechanical_reviewer` role briefing) for a parallel mechanical sub-pass on high-impact pieces. Never writes `status: greenlit` — that requires an explicit user signal (P8, R25). Use when the user says: "review this draft", "edit the article", "give me a review report on X", "is this draft ready", "check this content against the brief", "do an editorial pass", "evaluate this against the outline", or otherwise asks for editorial review of a draft. For greenlighting or publishing the result, route to `/content-writer` (revisions) or the orchestrator's publish routine (publish — there is no content-publisher skill) — the editor never transitions the draft itself.

Plans and executes research operations for a content project, evaluates source quality, deduplicates sources, extracts atomic facts with full provenance, and synthesizes findings into a research artifact downstream skills (content-strategist, content-writer, editor) can consume. Pairs with the `research-gatherer` sub-agent for bulk source ingestion, fact/quote extraction, credibility assessment, and pattern detection — research-analyst orchestrates; research-gatherer extracts. Use when the user says: "plan research on X", "gather sources for Y", "run research for this article", "synthesize the research findings", "what do we know about Z?", "extract facts from these sources", "build me a sources index", or otherwise asks for research planning, source ingestion, fact extraction, or synthesis on a content project. Writes to the `research/` Kind 2 zone at project root (NOT `workspace/research/`) per project-workspace-contract@2 (R62). Ends at status: draft → review only; never writes status: greenlit (P8 + R25 + R38 — orchestrator owns that mutation on explicit user signal).