newsroom-os

The campaign DESIGN engine of an installed newsroom — it takes a campaign the Managing Editor has already WARRANTED and designs it completely in ONE pass, forcing every dimension up front (shape, track structure & depth, per-piece audience job-to-be-done, source grounding, channel strategy, asset ladder, cadence, first-wave commission scaffold) instead of discovering missing dimensions one turn at a time. It emits a PROPOSED campaign master (Brief + Plan, source-grounded track backlog inline) at the not-yet-greenlit lifecycle state plus first-wave commission stubs, for the ME to review and the human to greenlight. It reads the running newsroom's OWN schema + reference docs and matches them — campaign shape, channels, and lifecycle vocabulary are never hardcoded. A Detection mode scans the corpus to SURFACE candidates for the ME's warrant test. Use when the user says "design the X campaign", "build out this campaign", "plan the <theme> program", "turn this cluster into a campaign", "scope the tracks/backlog for <campaign>", "campaign strategist", or (Detection) "scan for campaign candidates", "any campaign-shaped moments brewing?". Negative triggers: it does NOT decide the warrant (ME + human own that), write final commissions (stubs only), produce channel-native drafts (marketing handoff), or write per-article CCF briefs (content-strategist does, from the stubs), and NEVER sets active/greenlit/published or commits canonical state — it proposes, the ME commits.

Config-driven, channel-agnostic editorial-curation capability for an installed newsroom. The agent reads the source pool (filed/linked wire notes) and the user-approved Editorial Plan, clusters same-story signals semantically (Claude reasoning — NO vector search), checks prior coverage against the dossiers' coverage_history (dedup so the newsroom never re-litigates a dropped angle), applies the INSTALLED company-context relevance lens (Tier 1/2/3 read from config, never hardcoded), and writes ONE candidate shortlist artifact for the human editorial checkpoint. Two-layer reading: frontmatter-only for clustering, full content only for shortlisted clusters. Any relevance/priority signal it records is an INTERPRETED INSTRUMENT with a written interpretation — it NEVER composes into a verdict, NEVER auto-selects a story, and NEVER auto-writes a status:/arc_state:/plan_state: (R32/R68/P9). The ranked shortlist is a proposal; a human greenlights at the checkpoint (P8/R25) — the skill never auto-greenlights. Candidates are channel-agnostic (they may target any installed channel). Use when the user asks to "curate", "select topics", "what should we cover this issue", "review candidates", "run curation", "rank the source pool", "build the shortlist", or any request to turn the open source pool into a ranked editorial briefing. Manifest-first: reads MANIFEST.md before globbing zones.

Config-driven, channel-agnostic drafting engine for an installed newsroom. It turns approved research briefs + the commission + the user-approved editorial plan into a structurally-conformant edition DRAFT — in the channel's configured output_language, following the channel's configured format profile (section structure + toolkit), in the channel's configured voice lens. EVERY installation-specific value — the output language, the section names/order, the voice, the brand, the audience, the link/image rules — is a CONFIG READ from the active channel module's drafting.{format_profile_ref, voice_lens_ref, output_language} + config/company-context.md, NEVER a hardcoded literal. The drafting MECHANISM is pure: a reviewer grepping this skill finds no language/format/voice/brand/ section-name literals. Per D2 the voice lens is a PROPOSE→CONFIRM artifact the operator REFINES over several editions, and voice QUALITY is human-judged and explicitly NON-GATED; the skill gates STRUCTURE (format-profile conformance) + LANGUAGE (output_language) + PROVENANCE (every drafted claim grounds to a research brief's source anchor), not voice quality (AC4). The draft is written at status: draft and stops at the human refinement checkpoint — the skill NEVER self-greenlights and NEVER publishes (P8/R25). Use when the user asks to "draft the edition", "write the issue", "write the article", "assemble the draft", "rewrite a story", "produce the draft", or any request to produce or refine an edition draft after research briefs exist. Manifest-first: reads MANIFEST.md before globbing zones.

Markdown-driven validation + cache-regeneration capability for an installed newsroom. It walks the newsroom's primitive zones, validates each artifact's frontmatter against the bundled schemas, mechanically regenerates the canonical-vs-cache fields per the cardinality rules, runs the two-path provenance integrity audit over FROZEN edition snapshots, checks the reserved status-vocabulary separation (R61), and checks tag-vocabulary compliance — then writes a drift report. Manifest-first: it reads MANIFEST.md before globbing zones, and treats the meta/*-index caches as subordinate to the manifest (manifest wins on conflict). All JUDGMENT (tag normalization, provenance pass/fail interpretation, orphan/editorial-debt routing) stays in agent reasoning; the one bundled script does ONLY the mechanical cache recompute (E3 fence). Idempotent: a re-run on a clean newsroom produces a zero-drift report and no artifact writes. Use when the user asks to "audit the newsroom graph", "regenerate caches", "regenerate the meta indexes", "check provenance", "verify the provenance chain", "check canonical-vs-cache drift", "check status vocabulary", "check tag compliance", "find orphans", "graph audit", or any request about newsroom graph completeness. Also runs as a post-step after a distribution publish, as a preflight in the managing-editor briefing, and as the installer's P5 zero-drift dry-run baseline.

Config-driven, channel-agnostic INTAKE capability for an installed newsroom: it brings external signals in through the INGEST ADAPTER SLOT and emits one provenance-bearing WIRE NOTE per signal into newsroom/wire/. The agent DETECTs which ingest adapter the installation has wired (email-imap / m365-connector / rss-atom / http-api / manual-drop), fetches via that adapter, and DEGRADES GRACEFULLY when an adapter is absent — manual-drop is the irreducible floor that always works. Cleaning runs on the GENERIC CLEANER FLOOR (a defuddle-style pass that works on ANY source, lossy) as the v1 acceptance path; robust per-source cleaning is OPT-IN hand-authored refinement the operator adds later (PROPOSE→CONFIRM), NEVER auto-derived from a few samples (D1). Markdown-first (P13): ALL intelligence — what is signal vs noise, the signal_reason, the provenance anchor judgment — is the agent reading and writing prose; any bundled script is mechanical-only (fetch/convert/strip-tracking), never extraction or summarization. Wire notes conform to the bundled wire-note schema and carry a resolvable provenance anchor (url / corpus_note / raw). The skill never folds wires into dossiers (that is topic-editor) and never normalizes the controlled tag vocabulary. Use when the user asks to "ingest", "import sources", "new sources", "run ingestion", "scout the pool", "fetch the feeds", "pull the newsletters", or "bring these sources in". Manifest-first: reads MANIFEST.md before globbing zones.