human-writing git:20260813.da42b63A

Immutable. This exact content is served forever at /api/v1/blob/571a5409c967186c.

---
name: human-writing
description: "Use when drafting, rewriting, proofreading, diagnosing, or adapting source-grounded prose, especially technical articles, essays, tutorials, product notes, and social posts, while preserving facts, attribution, uncertainty, voice, and meaning. Supports private English-first reasoning for Chinese finals; not translation-only, fiction, or AI-detection evasion."
---

# Human Writing

## Outcome Contract

- **Outcome:** natural, source-grounded writing shaped for the requested reader and surface.
- **Evidence:** supplied text, notes, verified sources, author samples, and explicit task constraints.
- **Done when:** the artifact preserves meaning and protected details, solves the reader's task, and contains no concrete factual, structural, or voice defect worth another edit.
- **Output:** the requested artifact only, unless the user asks for diagnosis, alternatives, or editing notes.

## Core Stance

- Treat writing patterns as diagnostic signals, not a checklist or banned-word table.
- The author's supported voice wins over a generic house style.
- Over-editing is failure. Prefer fewer, stronger changes; leave already effective prose alone.
- Diagnostic severity does not expand edit permission. Respect the requested scope even when a broader rewrite might read better.
- Most useful polish is subtraction: remove repetition, empty framing, synthetic emphasis, and process narration without thinning the argument.
- Never trade factual precision, uncertainty, attribution, or a stable technical term for smoother prose.

## Workflow

1. **Lock the source.** Identify authoritative facts, judgments, unknowns, protected text, visibility limits, and claim status across the whole artifact. Before paraphrasing, make a private qualifier ledger for each claim that must be preserved or may be transformed, covering words or phrases that narrow frequency, priority, exclusivity, extent, lower bounds, confidence, or evidence status. When research is requested, also freeze its question, source boundary, and stop condition. Apply `references/fact-integrity.md` when the task contains multiple sources, technical claims, external facts, attribution, disclosures, or research-backed drafting.
2. **Set the target.** Infer the primary operation, reader, purpose, language, length, platform, genre, and permitted edit scope from the request and source. Treat an explicit minimum, maximum, range, or approximate length as an output contract: reserve a private section budget before drafting, using the requested unit rather than a substitute such as tokens or lines. Keep diagnosis depth, rewrite intensity, and edit scope separate. Ask only about missing information that would materially change the result.
3. **Calibrate voice.** Prefer a supplied author sample, then the stance and confidence visible in the source, then a neutral precise voice. Never manufacture first-person experience or quirks.
4. **Choose the smallest useful structure.** Use a private outline for new long-form or multi-claim work. For an existing long artifact, map sections and repetition before restructuring. Before writing a list, group the source by semantic dependency; sentence and clause boundaries are not item boundaries. One item must contain a rule and its boundary, a capability and the responsibility its limit leaves to the caller, a primary choice and the role of alternatives, or a stopping condition and its mandated next action. Keep independently triggered actions at different lifecycle stages separate. Skip planning ceremony for short or local edits.
5. **Draft or edit for substance.** Lead with the real question, scene, mechanism, or decision. Keep evidence near claims, integrate follow-up material where it changes the argument, and adapt for a platform rather than merely reformatting or translating.
6. **Run two passes.** First compare source and result for facts, actor-action-object relations, completion state, strength, protected text, disclosures, and safety. For every retained or transformed claim, check the qualifier ledger item by item; if the output has no unambiguous equivalent, restore the qualifier in the target language. When the user authorizes omission, audit the whole-claim omission instead of restoring an isolated qualifier. Do not delete a one-word qualifier from a retained claim merely because the sentence reads smoothly without it. Then diagnose clusters of template residue, weak logic, repetition, manufactured cadence, and voice drift. Fix only confirmed reader problems within the authorized scope. Finally measure every explicit length contract in its stated unit. If outside a range, revise and recount; do not estimate from screen length or stop merely because every planned section exists.
7. **Stop and return.** Compare with the source and stop when another change would only normalize preference-level wording. When evidence is missing, still return the safest useful part of the requested artifact when its supported structure and claims can stand on their own; use explicit placeholders only when the user requested a template or fill-ready draft. Return `Not enough context:` with the minimum missing facts only when no safe partial artifact can satisfy the reader's task. Never pad a minimum length with invented facts, repetition, or unsupported examples.

## Modes

- **Draft from source:** create prose from supplied or verified material without filling gaps with plausible fiction.
- **Rewrite or proofread:** improve a supplied artifact while preserving its position, meaning, structure, and protected content unless broader change was requested.
- **Diagnose:** identify concrete factual, logical, structural, voice, or platform problems without rewriting unless asked.
- **Platform adaptation:** reshape supported content for a named community or surface while preserving claims, attribution, disclosures, and author voice.

For a requested Chinese final from notes or mixed-language evidence, a private English claim-and-structure pass may help. Skip it for an existing Chinese draft or stronger Chinese sample unless requested. Treat the English intermediate as provisional phrasing, never as a new source, and check the Chinese final against the original evidence.

## Boundaries

Do not use for translation-only work, fiction or poetry, imitation of a living author's distinctive style, formal legal or regulatory writing, serious news reporting, technical correctness review as the primary task, unsupported advertising, or AI-detection evasion.

Hard requirements:

- Never invent experience, incidents, users, metrics, dates, versions, benchmarks, quotations, testimonials, results, rankings, or attribution.
- Preserve uncertainty, modality, claim status, actor role, and scope-bearing qualifiers such as `mainly`, `usually`, `often`, `only`, `at least`, priority, and degree. Preserve exact commands, paths, flags, identifiers, versions, code, and numeric ranges.
- Keep background research in the reasoning layer unless visible provenance is required or requested.
- Treat platform rules and other current external claims as volatile; verify them before presenting the artifact as compliant.
- Preserve frontmatter, links, code fences, quotations, and structural assets unless the task authorizes changing them.
- Integrate later instructions into the latest authoritative artifact; do not expose the editing sequence or leave append-only seams.
- Block only the affected unsafe or unsupported claim when a safe partial artifact remains possible.

## Output Contract

- Draft, rewrite, proofread, and adaptation: return the finished artifact only.
- Diagnose: return the excerpt or location, concrete problem, impact, and editing direction.
- Chinese final with private English support: return Chinese only unless both versions were requested.
- Unsupported artifact: return a safe supported partial artifact when useful; otherwise return `Not enough context:` followed by the minimum missing facts. A template may expose clearly marked evidence slots, but must not present them as completed claims.

## Reference Loading

- Load [fact-integrity.md](references/fact-integrity.md) for source precedence, claim status, attribution, disclosures, external claims, technical preservation, or a research-to-draft loop.
- Load [content-modes.md](references/content-modes.md) only when a genre changes the structure materially.
- Load [platform-calibration.md](references/platform-calibration.md) only for a named publishing surface; use profiles as heuristics, not current platform policy.
- Load [style-diagnostics.md](references/style-diagnostics.md) when voice, rhythm, Chinese technical prose, or AI-like template residue is the problem.
- Load [reasoning-and-explanation.md](references/reasoning-and-explanation.md) for mental-model repair, progressive explanation, boundaries, alternatives, or an earned conclusion.
- Load [revision-transparency.md](references/revision-transparency.md) only when an already-published artifact may need a visible correction or reader-relevant update note; routine content synchronization should be integrated naturally.
- Use [quality-rubric.md](references/quality-rubric.md) for final editorial review, [before-after-examples.md](references/before-after-examples.md) when a concrete calibration example helps, and [eval-cases.md](references/eval-cases.md) when changing Skill behavior.