Writing Core

Write for reader change, not for text production: a text is good when it asks the least cognitive work of its reader without losing meaning, uncertainty, or precision.

Hand off when the artifact has an owner, because each owner carries the golden path for that format: email, Slack, PO and executive summaries to writing-editorial; marketing email, hero, landing and web pages to writing-marketing; status, roadmaps, FaST, progress and product briefs to writing-product; Confluence and technical docs, PRs and commits to writing-engineering.

Frame before drafting

Reconstruct from the repository, ticket, diff, conversation, or documents: artifact, language, primary audience and what it already knows, intent, the reader's desired change (decide, understand, act), facts and evidence, real unknowns, tone, channel, and the terms this audience uses (code, API, docs, glossary), because a term the writer translates is one the reader must translate back. Ask only when a missing answer would change the artifact, because most gaps are recoverable from context and a question costs the user a round trip.

Never invent facts, customer evidence, metrics, decisions, test runs, deadlines, quotes, or certainty. Mark a gap instead, because one fabricated detail destroys the trust the rest of the text earned.

Workflow

1. Frame (above).
2. Shape the information architecture before sentences, because order decides whether the reader finds the point.
3. Draft from meaning: specificity over polish, verbs over abstractions, evidence over adjectives, one idea per sentence.
4. Edit in separate passes: structure (matches the reader's decision path?), meaning (every claim precise, uncertainty honest?), line (shorter, more concrete), language-native, copy (terms, spelling, dates), proof (names, numbers, links), then the AI-tell pass (references/ai-tells.md). Separate passes, because one combined pass fixes commas and misses a wrong order, and a line edit tightens a tell instead of removing it.
5. Verify: classify each claim as sourced fact, inference, recommendation, or unknown. Re-check time-sensitive or externally published facts at the source. For AI-assisted text, fact-check body and metadata by hand (references/verification.md).
6. Gate: from this skill's directory run python3 scripts/quality_gate.py --artifact <type> --lang <de|en> FILE with --artifact slides for deck text and --artifact prose for any other running text when no closer type is listed, and --facts holding the audience terms; fix every error, judge each warning.
7. Return the finished artifact, without commentary unless asked.

Language

Translation is meaning-preserving rewriting, never sentence-by-sentence substitution: rebuild meaning, relationship, intent, hierarchy, and action, then write natively in the target language.

• English: idiomatic US English unless the project says otherwise; active voice where responsibility matters; contractions in conversational text.
• German: current official orthography; concrete verbs over Nominalstil; one consistent form of address (du, Sie, or neutral); main point early; no literal English idioms or pseudo-English marketing syntax.

Details: references/de-style.md, references/en-style.md.

Editing priorities

Preserve in this order: truth, intended meaning, decision-relevant nuance, information architecture, audience fit, voice, brevity. Never shorten by deleting necessary uncertainty or evidence, because a short wrong text is worse than a long right one.

Readability scores are sensors, not objectives: a low Flesch or high LIX triggers a second look, never a rewrite that loses a needed term.

References

Load only what the task needs:

• references/ai-tells.md: the patterns that make text read as generated, the AI-tell pass, and the audience-term rule. Load for every draft.
• references/editorial-model.md: the FRAME to SHIP loop, edit passes, the universal golden paths per artifact family.
• references/de-style.md, references/en-style.md: house style per language.
• references/verification.md: claim classes, fact checks, test fixtures.
• references/quality-model.md: 0 to 4 rubric, gate thresholds, readability formulas and per-artifact gates.

Done when

The reader finds the point quickly, the required action or meaning is unambiguous, no unsupported fact was added, the language reads native, the structure fits artifact and audience, and remaining uncertainty is stated.

Maintaining these skills

Golden outputs in the domain references double as fixtures: each sits after a <!-- golden: <artifact> <de|en> --> marker, and tests/run_evals.py golden gates all of them. tests/test_quality_gate.py and tests/test_evals.py run in CI. tests/run_evals.py routing --host claude|codex replays tests/routing-cases.json in fresh headless sessions to show which skill a host actually loads; run it after changing any description.

references/sources.json is the evidence registry for all five writing skills (36 sources opened on 2026-10-21, each with access date, findings, limits, evidence level, and a revisit trigger); rules cite it as [SRC-WRITE-NNN], and the catalog's scripts/check-knowledge.py fails on a citation without an entry.