Human Writing

Overview

Turn supplied notes or drafts into source-grounded, natural writing without inventing experience or replacing the author's voice with a house style. human-writing names the intended result across drafting, rewriting, diagnosis, and platform adaptation; it is not an AI-detection-evasion or cosmetic humanizer tool.

Workflow

1. Lock the whole artifact. Build the source ledger across frontmatter, title, description, body, code, tables, links, and disclosures; then apply the precedence, provenance, semantic-fidelity, relationship, and disclosure rules in references/fact-integrity.md. Classify visibility separately from transformation policy. Treat profiles and examples as style-only. Do not write past the evidence.
2. Set claim state and actor role. Apply the canonical claim-status, provenance, and actor-role rules in references/fact-integrity.md. Do not infer decision authority from task execution.
3. Select the operation. Choose one primary operation and one primary genre per artifact. Allow a bounded secondary operation only when it directly supports the primary one, such as diagnosis before a requested rewrite. Split materially different language or platform outputs instead of blending their constraints. Treat follow-up additions as revisions to the latest authoritative draft unless the user requests a separate section or deliverable.
4. Set the target. Derive the reader, purpose, length, desired action, and evidence state from the supplied material before asking questions. Ask only for missing fields that would materially change the artifact; otherwise make low-risk editorial choices and proceed. Integrity and safety always win. Apply verified mandatory constraints before editorial preferences; use static platform profiles only as non-normative heuristics.
5. Calibrate the voice. Prefer the user's writing sample. Otherwise preserve the source's stance, confidence, person, and degree of involvement. Do not add first-person or experiential authority to neutral source material.
6. Choose the structure. Build around the real question, scene, decision, or task. For a new long-form or multi-claim artifact, make a private claim-and-evidence outline before prose; skip that ceremony for short-form and bounded local edits. When the reader needs a changed mental model, apply the smallest relevant tools from references/reasoning-and-explanation.md. The outline is planning, not the default deliverable. Do not default to 背景 / 现状 / 优势 / 总结 / 展望.
7. Draft or edit for substance. Convert the supported outline or source chain into finished prose. Delete empty framing, expose the actual judgment, connect claims to details, and preserve every distinct stage, role, condition, qualifier, and closing step in the supplied technical chain. Keep premises, derivations, judgments, intermediate models, and proxies distinguishable where material. Do not convert risks, principles, or future directions into lived incidents, completed transitions, or implemented guarantees. Integrate follow-up material where it changes the argument, then reread adjacent transitions, repetition, and the ending; do not merely append the latest instruction.
8. Run the human-writing pass. Audit the draft as the target reader. Detect clusters of template behavior, research-process leakage, editor commentary, diff-anchored narration in timeless documents, manufactured cadence, formulaic profundity, fake-candid openers, and unsupported specificity. Revise the affected passages rather than applying global punctuation, vocabulary, voice, or sentence-form bans. Preserve genuine habits, uncertainty, asymmetry, specific details, and protected secondhand text.
9. Run the integrity and safety pass. Apply references/fact-integrity.md and any applicable revision rules. Preserve attribution and uncertainty instead of repairing technical material from memory. Block only the affected claim or artifact when correctness is required, an action is destructive or irreversible, or an actual secret would be exposed; otherwise qualify the evidence state or route correctness review.
10. Classify acceptance. Use the four evidence layers in references/quality-rubric.md. Static validation or deterministic fixtures never prove real-model behavior or editorial acceptance.
11. Return the requested artifact or typed gap. Edit supported material when safe. Return Not enough context: with the minimum missing fields only when completing the requested artifact would require invention or unjustified certainty.