newsroom-drafting

newsroom-drafting

The drafting layer of the newsroom: it turns the approved, researched material for one commission into a structurally-conformant edition draft that goes to the editor-in-chief at the human refinement checkpoint. It is a skill, not a CLI and not a writer-of-record: the agent reads the inputs and writes the draft by applying editorial craft — but every installation-specific shape it follows (the language, the section structure, the voice, the brand, the link and image rules) is read from config at runtime, never assumed.

The make-or-break invariant — config-as-truth. This is the spec's flagged hardest genericization: "language/format/voice as config." A reviewer grepping this skill (and its references/checklists) for a language literal, a section name, a voice descriptor, a brand name, an audience description, or a banned-word list will find none. They are all resolved from the active channel module's drafting.{format_profile_ref, voice_lens_ref, output_language} and config/company-context.md. This is the literal generalization of the reference installation's output-lens port (one installation's lens replaced by another's, the lens system proven portable in the reference vault) — done once, generically, so the third installation costs an onboarding interview, not a fork.

This skill operates on an installed newsroom (a project-workspace-contract@2 workspace). It ships zero source-installation content.

Generic by construction. All worked examples below use the synthetic Meridian Robotics (a fictional industrial-automation vendor, German output) carried by the bundled contracts — mechanics only, never a real installation's language, sections, voice, or brand. The synthetic German fragments are illustrations of a config value resolving, not literals this skill embeds.

Where this sits in the spine (I/O contract)

Per the substrate I/O spine ([[newsroom-os/skills/newsroom-install/templates/substrate/contracts/io-contracts.md]]; drafting conforms to it and operates on the installed copy at runtime), drafting is the research + commission → edition-draft step that feeds the publish routine. It produces the content the distribution routine later freezes into an edition (per [[newsroom-os/skills/newsroom-install/templates/substrate/schemas/edition.md]], status: draft). Drafting does not itself publish, freeze the edition snapshot, or write delivery_state — that is the publish routine's job (R64).

approved research briefs ─┐
greenlit/draft commission ─┼─▶  DRAFTING  ─▶  story drafts + assembled edition draft  ─▶  human checkpoint (P8)
approved editorial plan ───┤                  (processing/stories/, processing/drafts/)        │
active channel drafting.* ─┘                                                                    ▼
config/company-context.md ─┘                                                  publish routine → edition (frozen, R64)

Direction	Artifact	Path	Role
Reads	registry	config/registry.yaml	resolve the active channel + its module_ref
Reads	channel module config	config/channels/<channel_id>.yaml	the drafting.{format_profile_ref, voice_lens_ref, output_language} triad
Reads	format profile	resolved from drafting.format_profile_ref	section structure + toolkit (config, never hardcoded)
Reads	voice lens	resolved from drafting.voice_lens_ref	voice/persona/vocabulary (iterated PROPOSE→CONFIRM artifact — D2)
Reads	company-context	config/company-context.md	§7 language & culture, §8 vocabulary, §4 audience, §1 identity
Reads	commission	newsroom/commissions/<id>.md	the production unit: brief_shape, derived_from_story_arcs, quick_note_items, slot plan
Reads	editorial plan	newsroom/editorial-plans/<plan-id>.md	format decision + slot directives the draft respects
Reads	research briefs	processing/research/<slot>/research-<slug>.md	the evidence boundary — source anchors per approved story
Reads (optional)	story-arcs / dossiers	newsroom/story-arcs/, newsroom/dossiers/	arc angle + dossier evidence backing
Writes	story drafts	processing/stories/<slot>/story-<slug>.md	one self-contained story block per full-story slot
Writes	assembled edition draft	processing/drafts/<edition-draft-id>.md	the issue/article draft → human checkpoint (status: draft)

processing/** is operational, not a primitive zone. Drafting's output lands in processing/stories/ and processing/drafts/ — in-flight editorial workspace, not one of the seven newsroom/ primitive kinds (directory-structure.md). It carries no newsroom-index manifest entry and there is no meta/drafts-index; the draft zone is never manifest-indexed or graph-audit-regenerated. The publish routine reads the assembled draft directly to produce the newsroom/editions/<id> edition.

Operational runbook. Read references/drafting-runbook.md before executing — it carries the step-by-step procedure (config resolution, per-story drafting from briefs, the format-profile-driven structure, the voice-lens application as iterated-non-gated, assembly, and refinement). The format-profile schema lives at references/format-profile-schema.md; the voice-lens schema at references/voice-lens-schema.md; the AI-authenticity rubric (per-language machine-tell catalog + humanization moves, applied on top of the voice lens) at references/ai-authenticity/. The end-of-run gates live in checklists/structural-conformance.md and checklists/provenance-grounding.md.

Phase 0 — manifest-first (R62)

Read MANIFEST.md at the newsroom root FIRST, before any files-first globbing of zone folders. The manifest is the routing surface: it lists which commissions, editorial-plans, story-arcs, and dossiers exist with their kind / status / path / edges. Discover the commission to draft and its upstream artifacts via the manifest; use the zone folders only to resolve the concrete files the manifest points at. The meta/*-index caches are derived caches subordinate to the manifest (manifest wins on conflict). Files-first discovery — globbing a zone to learn what exists — is the refused anti-pattern.

Phase 0.5 — resolve the active channel's drafting config (the config-as-truth gate)

Before writing one word, resolve the active channel module and its drafting config. This is the step that makes everything downstream config-driven:

1. Find the active channel. The commission carries target_channel. Read config/registry.yaml; find the channels[] entry whose channel_id maps to that target_channel and is enabled: true; load its module_ref (config/channels/<channel_id>.yaml).
2. Read the drafting triad from that module config:
   • drafting.output_language — the ISO 639-1 language the draft is written in.
   • drafting.format_profile_ref — resolve to the format profile (section structure + toolkit). See references/format-profile-schema.md.
   • drafting.voice_lens_ref — resolve to the voice lens (voice, persona, vocabulary, framing discipline). See references/voice-lens-schema.md.
3. Validate the *_ref paths resolve INSIDE the install (config/... or a bundled child under the active module skill's references/). A *_ref that escapes the install dir (../, an absolute system path, a sibling-plugin path) is invalid — STOP and report (R63 / §3.5 isolation at the config layer).
4. Read config/company-context.md — §7 (language & culture), §8 (vocabulary: brand spelling rule + words-to-use + words-to-avoid), §4 (audience), §1 (identity). These supply the installation values the format profile and voice lens reference; the skill reads them, it does not assume them.

What this means in practice (the literal generalization). The reference installation hardcoded "write in German", a fixed Was passiert ist / Einfach erklärt / Quellen section set, a named output lens, and a brand-spelling scan. Here, all four of those are reads: language ← drafting.output_language; sections ← the format profile; voice ← the voice lens; brand spelling ← company context §8. A different installation supplies a different language, a different section set, a different voice, a different brand — and this skill produces a correct draft for it with no skill edit. That is the portability claim, and the structural-conformance checklist is what enforces it.

If the active channel cannot be resolved, or any required drafting.* field is missing, or a *_ref does not resolve inside the install — STOP and surface the gap. Drafting never falls back to a hardcoded default language/format/voice (there is none to fall back to — that is the point).

Phase 0.6 — the input gate (P8 / R25)

Condition	Action
No greenlit/draft commission for this slot	STOP. Tell the user to run the managing-editor/commission step first.
Commission exists but a derived_from_story_arcs story has no research brief	STOP. Send that story back to story-research. Drafting never invents evidence.
Research brief quality_status: warnings and lacks support for a central claim	STOP. Return to research rather than drafting from weak evidence.
An approved editorial plan exists for the slot	Read it; respect its format_decision + slot directives. Record the linkage in the draft frontmatter.
A plan_state: proposed plan exists	STOP. Surface it — a proposed plan awaits the user's approval; do not silently bypass the gate.

Drafting is the executor of an already-approved commission; it does not make the selection or planning decisions (those belong to curation + the managing-editor + the human checkpoint). It reads status:/plan_state: to know it is clear to run; it never writes a status:/arc_state:/plan_state:/dossier_state: transition (R32 / R68 / P9) and never writes status: greenlit on anything (P8 / R25).

The drafting mechanism (PURE — all judgment is the agent's; all specifics are config)

Four moves, in order. Moves 1–3 produce the story drafts and the assembly; move 4 hands the assembled draft to the human checkpoint. None of them transitions a status:/*_state: autonomously, and none of them embeds an installation literal.

Move 1 — Per-story drafting from the brief (the format profile drives the shape)

For each full-story slot the commission carries (derived_from_story_arcs + the plan's slot breakdown), write one self-contained story draft to processing/stories/<slot>/story-<slug>.md:

• The evidence boundary is the research brief. Draft only from the brief's primary_source_anchors / secondary_source_anchors and body. Drafting is closed to broad web research for prose claims — if the brief is insufficient, STOP and return the story to story-research (the format profile's IMAGE rule, if it declares one, is the only narrowly-scoped carve-out — see Move 1 image step).
• The story SHAPE comes from the format profile, not from this skill. The format profile declares the channel's section toolkit (the available story beats), the mandatory minimum (which beats every story must carry), the count discipline (how many beats per story), the section labelling style (e.g. bold inline labels vs subheaders — a config value, not a literal here), and the default order. Read those from the resolved format profile and apply them. Do not assume any particular section name exists — read the toolkit.
• The LANGUAGE is drafting.output_language. Write the prose in that language. Apply the company-context §7 orthography/register notes for that language. (The reference installation's German-vs-ASCII-umlaut discipline generalizes to "respect the configured language's orthography," resolved from §7 — not a hardcoded German rule.)
• The VOICE is the voice lens. Apply the resolved voice lens's persona, vocabulary (words-to-use / words-to-avoid from the lens + company-context §8), framing discipline, and relevance-angle mechanism. Voice quality is human-judged and non-gated (D2) — apply the lens as best you can; the operator refines it over editions. (See "D2 — the voice lens is iterated; voice quality is non-gated.")
• Authenticity over machine-tells. On top of the voice lens, write so the prose does not read as machine-made. Consult references/ai-authenticity/ (00-method.md
  • the en/de/nl catalog matching drafting.output_language) and apply its humanization moves: vary sentence length, cut formulaic signposting, prefer concrete specifics over generic abstractions, break the "not just X, it's Y" and rule-of-three reflexes, drop hedge-openers, and for DE/NL restore the modal particles a machine drops. Like voice quality, this is human-judged and non-gated (D2 / R32) — a craft aid, never a score or a blocker.
• Event-first construction. Open with the external event (what was published / announced / measured, by whom, the key verified finding) before interpretation — this is a channel-agnostic craft rule the format profile may reinforce. Do not open with the editorial conclusion.
• Provenance grounding (AC10 chain). Every retained claim must trace to a research-brief anchor. The story draft's source_anchors: frontmatter mirrors the brief's anchors so the post-publish provenance integrity audit can walk the full-story path (Edition → Commission → Story-Arc → Dossier → Wire → anchor). A claim with no brief anchor is a provenance dead-end — cut it or send the story back to research. See checklists/provenance-grounding.md.
• The IMAGE step (if the format profile declares an image affordance). Many channels want a content image per story. If the format profile's toolkit declares an image slot, run the format profile's image rule (typically: scan the source package + the brief's primary-source anchors for a content image — chart, screenshot, diagram, product visual — never chrome/logos/tracking pixels; attribute to the original creator, not an aggregator). Record the image decision in story frontmatter (has_image + an image_scan_log when none found). If the format profile declares no image affordance, skip this step entirely. The image rule is read from the format profile, never hardcoded here.

Move 2 — Draft the channel-level synthesis slot (only if the format profile + plan declare one)

Some channels' format profiles declare a channel-level synthesis/lead slot (a cross-story editorial piece). Draft it only when the format profile defines it and the editorial plan approved it. It follows the format profile's spec for that slot (length, image budget, free-form vs toolkit), written in drafting.output_language and the voice lens. Save to processing/stories/<slot>/<synthesis-slug>.md. If the format profile declares no such slot, this move is a no-op.

Move 3 — Assemble the edition draft (the format profile drives the issue structure)

Combine the story drafts + the commission's quick_note_items into one assembled edition draft at processing/drafts/<edition-draft-id>.md:

• The ISSUE structure comes from the format profile — the header/title form, the channel-level section order (synthesis slot placement, story-block order, quick-note section placement, any tool/extra sections), the quick-note bullet format, the link-resolution rules (e.g. "no tracking redirects; resolve to the canonical destination"), and the footer/signature form. All read from the format profile + company-context §8 (brand), never hardcoded.
• Story drafts are the source of truth for story prose (and their image+caption blocks). Assembly may reorder, shape the issue headline, and phrase the quick-note section — but may not rewrite story content without writing the change back to the story draft first (the canonical-source rule — see Move 4 / refinement).
• Quick notes ride the commission, not a story-arc. Each quick_note_items[] entry has ≥1 wire_refs (the provenance hard rule). Render each per the format profile's quick-note format, carrying a resolvable link to the destination anchor. Their provenance resolves along the quick-note path (Edition → included_quick_note → Wire → anchor) — keep each note's wire_refs intact so the publish routine can freeze them into the edition's included_quick_notes[].
• Brand spelling. Apply company-context §8's brand-spelling rule on a final pass (the generalization of the reference installation's case-sensitive brand-variant scan — the rule is read from §8, the scan is the mechanism).
• The assembled draft is status: draft. Record output_language, target_channel, commission_ref, story_drafts, story_order, and the plan linkage in frontmatter (schema below). It is not an edition yet — the publish routine produces newsroom/editions/<id> from it.

Move 4 — Hand the draft to the human refinement checkpoint (P8 / R25)

Present the assembled draft to the editor-in-chief with a brief summary (story count, length against the format profile's target, any trade-offs made). Refinement is human-in-the-loop:

• Accept feedback (conversation, inline comments, or a feedback file).
• The canonical-source rule. If feedback changes story content, update the processing/stories/<slot>/story-<slug>.md story draft first, then reassemble. The assembled draft must never become the only place where the final story text lives. Issue-level-only changes (story order, header wording, quick-note phrasing, transitions around the synthesis slot) may be edited directly in the assembled draft.
• Drafting never writes status: greenlit, never sets delivery_state, and never publishes. The greenlight is the user's explicit signal at the checkpoint (P8 / R25); publishing is the orchestrator's publish routine + the project-local adapter (R64). Drafting stops at a clean status: draft the human can approve.

D2 — the voice lens is iterated; voice quality is non-gated (the honest ceiling)

The hardest part of "voice as config" is honest about its own limit:

• The voice lens is a PROPOSE→CONFIRM artifact the operator refines over several editions (D2). It does not arrive publication-perfect from the onboarding interview. Drafting applies the current lens; the operator sharpens it issue over issue, and a sharper lens shows up in later drafts with no skill change.
• Voice QUALITY is human-judged and explicitly NON-GATED (R32 / R68 / P9). There is no composite "is this good writing" score, and the skill never blocks, refuses, or auto-revises a draft on a voice-quality verdict. A reviewer should not expect a fresh install to produce publication-grade voice on edition 1; the design does not promise it.
• What IS gated (AC4): STRUCTURE (the draft conforms to the format profile's section structure + minimums + order), LANGUAGE (the draft is in drafting.output_language, not the source-installation's language), and PROVENANCE (every claim grounds to a research-brief anchor — AC10 chain). AC4 passes on structural conformance + correct target language + zero source-installation residue, not on a voice score.

This split is load-bearing: it is why "voice as config" is an honest claim and not an over-promise. See references/voice-lens-schema.md § "Voice quality is non-gated."

Generic quality gates (format-agnostic / config-driven)

These are the mechanisms, all driven by config — none embeds an installation literal:

1. Structural-conformance gate (STRUCTURE — gated). The assembled draft conforms to the resolved format profile: every story carries the mandatory minimum beats, the per-story beat count is within the profile's discipline, the labelling style matches, the issue-level section order matches, and the synthesis /quick-note slots appear iff the profile + plan declare them. Checklist: checklists/structural-conformance.md.
2. Language gate (LANGUAGE — gated). The prose is in drafting.output_language (proper nouns and unavoidable technical terms excepted), and respects that language's orthography per company-context §7. Brand spelling matches §8.
3. Provenance-grounding gate (PROVENANCE — gated, AC10). Every retained claim traces to a research-brief anchor; every quick note keeps ≥1 wire_refs; the story drafts' source_anchors: mirror the briefs so the post-publish provenance integrity audit can walk the typed paths. Checklist: checklists/provenance-grounding.md.
4. Image gate (if the format profile declares an image affordance — mechanical). The image decision is documented in every story draft's frontmatter (has_image
   • image_scan_log when none); the assembled draft preserves every story image + caption verbatim. If the format profile declares no image affordance, this gate is N/A.
5. Link-resolution gate (config-driven). Outbound links obey the format profile's link rules (typically: resolve every tracking redirect to its canonical destination; strip tracking params). The format profile names the rule; the skill applies it.
6. The canonical-source rule (refinement). Story drafts remain the canonical home of story prose; the assembled draft is never the only place a final story lives.

Voice quality is NOT on this list — by design (D2). None of these gates scores or blocks on voice quality. Voice is applied (the lens) and human-judged (non-gated).

Frontmatter schema — story draft (processing/stories/<slot>/)

---
type: story-draft                            # processing artifact — `type:` only; `kind:` is reserved for the seven manifest primitives
story_kind: full-story
story_slug: <arc-slug>                       # mirrors the story-arc arc_id + research brief story_slug
title: "<plain working title>"
target_channel: <from commission>            # closed enum value — read, not hardcoded
output_language: <from drafting.output_language>   # ISO 639-1 — read from channel config
issue_slot: "<period>"                       # mirrors the editorial-plan / commission slot
commission_ref: "[[newsroom/commissions/<commission-id>]]"
research_brief_ref: "[[processing/research/<slot>/research-<slug>]]"
story_arc_ref: "[[newsroom/story-arcs/<arc-slug>]]"
format_profile_ref: "<resolved drafting.format_profile_ref>"   # provenance of the shape applied
voice_lens_ref: "<resolved drafting.voice_lens_ref>"           # provenance of the voice applied
selected_sections: []                        # the beats chosen FROM the format profile's toolkit
source_anchors:                              # MIRRORS the research brief — the AC10 grounding
  - anchor_type: url | corpus_note | raw
    ref: "<https://… | [[knowledge/<subzone>/<note>]] | [[research/sources/<id>/raw…]]>"
    backing_wire: "[[newsroom/wire/<wire-id>]]"
has_image: false | true                      # only when the format profile declares an image slot
image_scan_log: null                         # required when has_image is false AND an image slot exists
status: draft                                # RESERVED enum (R61); starts + stays draft (never greenlit here)
created: 2026-06-18
related:
  - "[[processing/research/<slot>/research-<slug>]]"
  - "[[newsroom/story-arcs/<arc-slug>]]"
---

Frontmatter schema — assembled edition draft (processing/drafts/)

---
type: edition-draft                          # processing artifact — `type:` only; `kind:` is reserved for the seven manifest primitives
edition_draft_id: "<channel slot id>"        # mirrors the channel's canonical slot/filename pattern
title: "<published-style title>"             # in output_language; brand spelling per company-context §8
target_channel: <from commission>            # closed enum — read, not hardcoded
output_language: <from drafting.output_language>   # ISO 639-1 — read from channel config
issue_slot: null | "<period>"                # for issue-based channels
commission_ref: "[[newsroom/commissions/<commission-id>]]"
editorial_plan_ref: null | "[[newsroom/editorial-plans/<plan-id>]]"
format_profile_ref: "<resolved drafting.format_profile_ref>"
voice_lens_ref: "<resolved drafting.voice_lens_ref>"
format_decision: null | regular | special | series | digest   # mirrors the plan
story_drafts:
  - "[[processing/stories/<slot>/story-<slug>]]"
story_order:
  - <arc-slug>
synthesis_slot_included: false | true        # iff the format profile + plan declare one
quick_note_slugs: []                         # mirrors commission.quick_note_items[].slug
status: draft                                # RESERVED enum (R61); the human greenlights, not this skill
created: 2026-06-18
last_updated: 2026-06-18
related:
  - "[[newsroom/commissions/<commission-id>]]"
  - "<resolved drafting.format_profile_ref>"
  - "<resolved drafting.voice_lens_ref>"
---

The edition-draft is a workspace artifact, not the published edition. The publish routine consumes it to write newsroom/editions/<id> (frozen snapshot + delivery_state) per the edition schema (R64). Drafting writes neither the frozen included_story_arcs[]/included_quick_notes[] snapshot nor delivery_state — those are the publish routine's (the truth-gate is publish's, not drafting's).

Error handling

Condition	Action
Active channel not resolvable from registry.yaml / commission target_channel	STOP. Report the registry/commission mismatch; never guess a channel.
A required drafting.* field missing, or a *_ref escapes the install	STOP. Report the invalid module config (R63). No hardcoded fallback exists.
Format profile or voice lens does not resolve	STOP. Report the broken *_ref. Drafting cannot run without a format profile + lens.
Commission missing or not greenlit/draft	STOP. Route back to the commission step.
Research brief missing for a derived_from_story_arcs story	STOP. Send that story back to story-research.
Research brief stale or quality_status: warnings without central-claim support	STOP. Regenerate/strengthen research before drafting.
A drafted claim has no research-brief anchor	Cut the claim or return the story to research — never publish an ungrounded claim.
quick_note_items[i].wire_refs empty	STOP. A quick note with no wire is a provenance dead-end (Commandment V).
No quick_note_items	Proceed without the quick-note section (if the format profile makes it optional).
Format profile declares no synthesis slot, but the plan asked for one	Surface the mismatch; the format profile is authoritative for what slots exist.

Acceptance criteria

• Produces, for an approved commission, story drafts + one assembled edition draft at status: draft.
• Config-as-truth: a grep of this skill (+ references/checklists) finds no hardcoded language/format/voice/brand/section-name literal; all are config reads (Phase 0.5 + the format-profile + voice-lens schemas).
• AC4: the draft is structurally conformant to the format profile AND in drafting.output_language AND free of source-installation residue — gated on structure + language + provenance, not on a voice-quality score.
• D2: the voice lens is applied as an iterated artifact; voice quality is human-judged and non-gated.
• Provenance (AC10): every drafted claim grounds to a research-brief anchor; story drafts mirror anchors; quick notes keep ≥1 wire_refs.
• P8 / R25: the skill never writes status: greenlit and never publishes; it stops at the human refinement checkpoint.
• Conforms to the substrate I/O contracts + the edition schema; the draft is a workspace artifact the publish routine turns into an edition.
• R63 / P13: self-contained; markdown-first (no script does any judgment work).

Language Handling (P12)

The draft's human-facing output language is config-driven, never hardcoded — it is drafting.output_language (the active channel module's triad), which the install derives from config/company-context.md §7 output_language. Respond to the operator in the conversation's language unless asked otherwise. Beyond that:

• Detect the conversation language; respond in the user's language unless explicitly asked otherwise.
• The drafted edition is written in drafting.output_language (Phase 0.5) — honor that config declaration over the inferred conversation language for the artifact prose; it is the gated LANGUAGE invariant (AC4), not a default.
• Translate internal checklist labels and section headings naturally into the output language — do not force English headings into non-English drafts.
• Preserve legal, financial, and technical terms in their original language; respect the configured language's orthography per company-context §7.
• Preserve Unicode characters natively (ü, ö, ä, ß, etc.); never substitute transliterations (ue, oe, ae, ss).
• Frontmatter field names and enum values stay English regardless of content language (type, status, target_channel, output_language value as ISO 639-1).

End-of-run — self-improvement (P15)

Standard runs end normally: the assembled status: draft edition draft + a brief summary + the suggested next step (the human refinement checkpoint). Do not fire a generic "anything to add?" prompt.

When a deviation specific to drafting occurred during the run, surface it and ask whether to fold the change back into SKILL.md / the bundled references (the format- profile schema, the voice-lens schema, the structural-conformance / provenance- grounding checklists) before going idle. Drafting-specific triggers:

• the operator rewrote story prose at the checkpoint in a way that reveals a missing beat or labelling convention in the resolved format profile;
• the voice lens missed the mark and the operator's correction implies a sharper lens directive (D2 — propose folding it into the voice lens, not the skill);
• a format-profile or voice-lens *_ref resolved awkwardly, or an image/link rule needed improvisation the profile did not anticipate;
• a provenance-grounding or structural-conformance gate fired in a way that suggests the checklist wording is unclear or incomplete.

Name the concrete artifact (the specific story draft, the format profile, the voice lens) in the prompt. If nothing deviated, end without the prompt.