# Document Writing > Write, rewrite, or audit truthful prose for a real document, reader, and purpose without reducing quality to an anti-AI phrase list. Use for READMEs, design and decision documents, reports, explainers, guides, essays, internal notes, public project copy, or requests such as "write this document", "rewrite this README", "make this sound less AI-generated", "好好表达", "去 AI 味", "改写文档", and "这段话写得不像人". Do not use for choosing a shared name, visual layout, literal transcription, or mechanical formatting. - Skill: `lidessen/document-writing` (Agent Skill, multi-file: 4 files) - Install (CLI): `npx skillmds@latest add lidessen/document-writing` - Raw SKILL.md: https://api.skillmd.com/api/skills/lidessen/document-writing/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Security - Author: lidessen (https://skillmd.com/u/lidessen) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/lidessen/document-writing --- # Document Writing ## Principle expression **Primary:** P16 **Supporting:** P01, P05, P07 ## Scope Own one judgment: **how should source-grounded meaning become a document that this writer can truthfully address to this reader, in this situation, so the reader can understand, decide, or act?** Treat “AI flavor” as an observed failure, not a genre or a banned-word problem. Recover the missing relation before editing its symptoms. A list, heading, short sentence, formal word, or em dash may be right in one document and wrong in another. Use a host Sequence and matching interpretations when present. Otherwise use the read-only fallback in `references/sequence.md`. Read only P16, P01, P05, and P07 for this skill's stable lineage. ## Start Recover the smallest writing situation that changes the draft: ```text Deliverable and authorized edit depth: Source facts, claims, and protected wording: Writer and accountable point of view: Reader, prior knowledge, and relation to the writer: Why this document exists now: Change it should produce in the reader: Medium, genre, length, language, and publication constraints: Closest useful voice sample, if the user supplied one: ``` Infer low-risk details from the actual artifact and request. Ask only when a missing answer would materially change the document. For a short local edit, do not force a writing brief or expose this packet to the user. Read `references/concepts.md` only when the writing relation or movement is unclear, sources must be reconstituted across several documents, a structural rewrite is required, “AI味” has survived an ordinary revision, or voice is disputed. Skip it when the source, reader, purpose, and document shape are already explicit, even if the finished document is long. ## Core method 1. **Bound what may be said.** Read the full source needed for the requested scope. Before drafting, classify the relevant material internally as publishable proposition, attributed judgment, constraint-only instruction, or unknown. A prohibition constrains the draft; it is not itself a fact to print. Map each reader question the brief requires to a publishable source proposition. Preserve negation, modality, scope, causality, chronology, attribution, uncertainty, names, numbers, citations, and defined terms. Preserve absences too: do not complete a missing year, actor, unit, cause, motive, comparison, or project state. Compare the requested reader questions with the source: if the user asks for “why,” a benefit, an assurance, or a conclusion the source cannot establish, do not answer it from temporal adjacency or a plausible project intention. Use the nearest supported relation, disclose the gap, or ask when it changes the deliverable. Let the density of distinct publishable propositions and relations constrain the supported scope and length. Explanation may unfold a supported relation, but paraphrase, repetition, and plausible extrapolation do not create more material. When the requested length exceeds what the source can sustain, research or ask only when authorized; otherwise narrow the question or deliver a shorter truthful document. 2. **Recover the writing relation.** Name who is speaking to whom, why now, and what should be different after reading. Replace the generic assistant-to-any-user relation with the artifact's actual relation. This decides register, explanation depth, stance, courtesy, and whether the reader needs prose, headings, a table, or a list. 3. **Find the document's movement.** State the controlling relation in one sentence, then choose a spine the material can support: answer to reason, concrete observation to explanation to consequence, problem to mechanism to decision, chronology, comparison, or procedure. Do not create independent topic buckets merely because the source contains several categories. 4. **Draft positively.** Write from the source boundary, relation, and movement. Let paragraphs depend on one another through cause, contrast, reference, qualification, example, or consequence. Use the reader's needed terms. When voice matching was requested, prefer explicit user guidance and the closest same-medium samples; imitate relevant choices, not quirks or a universal personality profile. 5. **Revise in causal order.** - **Truth:** compare each material claim and implication with its source. - **Whole:** verify that the opening, movement, emphasis, and ending produce the intended reader change. - **Paragraph:** identify the new proposition, relation, distinction, action, or consequence each paragraph contributes. Merge, delete, or reorder paragraphs that only restate earlier material; remove catalog sections, repeated premises, fake transitions, and freely reorderable blocks. When the draft appears inflated, remove a substantial passage and compare the result. Keep the compression when source-supported meaning, necessary qualification, and the reader's next action remain intact. - **Sentence:** restore actors, exact verbs, warranted confidence, reference, and varied syntax where the relation requires it. - **Surface:** only now inspect clusters of model habits such as excessive headings, list-shaped prose, repeated contrast formulas, slogans, ceremony, uniform cadence, or unsupported friendliness. 6. **Deliver the document, not the workshop.** Edit the target file when authorized. Otherwise return publication-ready prose. Include an audit, rationale, or alternatives only when requested or when an unresolved source conflict prevents a truthful final document. ## Model and voice boundary Treat the model label as a clue, never as the diagnosis. Observe the actual draft: one execution may over-format, another may ignore the requested language, and another may create persuasive but unsupported connective claims. Correct the observed failure without growing a universal blacklist. Do not build a persistent personal voice profile unless repeated cross-document work and user evidence establish that need. One supplied sample licenses task-local imitation only. Never fabricate roughness, inconsistency, opinions, anecdotes, or self-correction to simulate a human author. ## Boundaries - Preserve information while changing expression; do not silently turn `may` into `will`, association into causation, or an attributed claim into fact. - Do not ban a surface form globally. Use pattern clusters as late diagnostic evidence and keep forms that serve the medium. - Do not turn every document into continuous prose. Reference material, procedures, release notes, and comparison tasks may genuinely require independently scannable structure. - Route shared naming and operative definitions to `naming-and-articulation`; route visual hierarchy and document presentation to `visual-design`. - Do not claim that polished prose is human-authored or that a detector score establishes authorship. ## Completion standard The document is ready when its material claims remain source-supported; its writer, reader, situation, and intended change are legible in the choices it makes; its parts form one useful movement rather than a catalog; its medium supports the reader's action; and a final pass finds no dominant accidental model pattern. A semantic reviewer may judge whether the expression is right, but only the authorized human or owning process accepts consequential content.