| name | write-documentation |
| description | Use when writing or substantially rewriting human-facing prose: documentation, README, guides, blog posts, emails, Slack messages, PR descriptions, release notes, or any text a human will read. Not for code comments, commit messages, or agent-to-agent communication. |
Write Documentation
Announce at start: "I'm using the write-documentation skill to write human-quality prose."
Purpose
Write for the actual context.
The goal is prose that fits the medium, the task, and the reader. If it does that well, it will usually read as human-authored as a side effect. Do not optimize for "sounding human." Do not optimize for beating detectors. Both produce worse writing.
When to Use
- Writing or substantially rewriting any human-facing text
- Creating documentation from scratch (README, guides, tutorials, API docs)
- Drafting blog posts, articles, emails, or Slack messages
- Rewriting a section flagged by
document-release as needing major revision
- User asks for help with prose quality, tone, or style
- Writing PR descriptions, release notes, or CHANGELOG entries
When NOT to Use
- Code comments and docstrings (for developers reading code)
- Commit messages (their own conventions)
- Agent prompts, skill files, CLAUDE.md (for AI agents, not human readers)
- Minor edits: fixing a typo, updating a version number, adding a link
document-release work that is factual syncing, not prose rewriting
Beads Integration
bd create "Write: <description of what's being written>" -t task
Closed with evidence at Step 6 — the single source of truth for the close command.
Core Workflow
digraph write_documentation {
rankdir=TB;
node [shape=box, style=rounded];
context [label="Step 1: Identify Context\nMedium, audience, reader need, register"];
structure [label="Step 2: Structure Decision\nTask-oriented → answer first\nLong-form → through-line + example"];
draft [label="Step 3: Draft\nRules internalized, not mechanical\nFocus on substance"];
checks [label="Step 4: Required Checks\nShort: 1-3, 5, 7, 10\nLong: all 1-10"];
cut [label="Step 5: Cut\nGeneric, ceremonial, over-engineered"];
present [label="Step 6: Present to User"];
revise [label="User requests changes?" shape=diamond];
done [label="Close bead with evidence" shape=doublecircle];
context -> structure -> draft -> checks -> cut -> present -> revise;
revise -> draft [label="yes"];
revise -> done [label="no"];
}
Step 1: Identify Context
Before writing anything, answer these four questions:
- What is this text? (medium: README, email, blog post, Slack message, guide, PR description...)
- Who reads it? (developers, end users, managers, the public...)
- What do they need? (an answer, a next action, understanding, a decision...)
- What register? (formal technical docs, casual chat, professional email, marketing...)
Done when: all four questions have explicit answers.
Step 2: Structure Decision
- Task-oriented text: Identify the answer or next action that belongs first.
- Long-form text: Decide the through-line and one concrete example, moment, or case that can carry real weight in the piece.
- Apply medium routing (see below).
Done when: the answer/next-action (task-oriented) or through-line-plus-example (long-form) is decided, and medium routing is applied.
Step 3: Draft
Draft to fit the identified context, not an abstract idea of "good writing." The core rules below should be internalized during drafting, not mechanically applied paragraph by paragraph. Focus on substance over style.
Done when: a complete draft exists covering the identified context.
Step 4: Required Checks
Run the revision checks scaled to the length and stakes of the piece. Fix issues inline during this pass. See the Required Checks section below.
Done when: the scaled check set has been run and every issue it found is fixed inline.
Step 5: Cut
Cut what sounds generic, ceremonial, over-engineered, suspiciously over-specific, or too cleanly modular. Collapse paragraphs that restate each other. Replace the most generic clause in the piece with something specific or delete it.
Done when: the piece has been reviewed for generic/ceremonial language and the most generic clause is fixed or gone.
Step 6: Present to User
Show the final text. If the user requests changes, loop back to Step 3. When approved, close the bead: bd close <id> --reason "Written: <what>, checks run: <which checks passed>".
Done when: the user has approved the text and the bead is closed with evidence of which checks ran.
Precedence
When rules conflict:
- Truth, safety, accessibility, and platform/legal requirements
- Explicit user instructions
- Genre and medium norms
- Core rules
- Optional watchlists and heuristics
If the user asks for bullets, use bullets. If accessibility or the medium require structure, use structure. If the user asks for a neutral summary, do not force first person or extra stance into it.
Medium Routing
- Chat, comments, replies, DMs, forum posts: Running prose by default. Use lists only when the information is naturally list-like or the user asked for one. Avoid decorative formatting and canned support tone. In plain-text contexts, prefer straight ASCII quotes and apostrophes. Prefer commas, colons, conjunctions, subordinate clauses, or full stops over em dashes unless the dash clearly earns its keep. Do not replace every dash with a period; if the second thought is still part of the first turn, keep the sentence moving.
- Email between colleagues: Usually prose first; lists are fine for discrete items, decisions, or action points.
- Documents, specs, reports, technical writing: Structure is expected. Use headings, bullets, and sequence when they help scanning and precision.
- Web pages, help centers, UI text, and public docs: Put the answer or next action early. Preserve scannability and accessibility: descriptive headings, lists for steps, descriptive link text, and plain alt text when images carry information. Do not flatten useful structure just to avoid looking templated.
- Blog posts, articles, criticism, retrospectives: Use structure on purpose. Pick an angle. Do not let dates, named milestones, or neat category buckets become the spine unless the user explicitly asked for that structure.
- PR descriptions, release notes, CHANGELOG entries: Lead with what changed and why. Structure for scanning. Concrete consequences over ceremonial phrasing.
Safety Rails
These are not AI tells by themselves: em dashes, semicolons, however, competent punctuation, well-formed paragraphs, and the right word even if it appears on somebody's banned list.
Do not invent typos. Do not break grammar on purpose. Do not inject slang, profanity, fake uncertainty, or staged messiness to simulate humanity. No mandatory actually turn. No manufactured negativity. No programmatic sentence-length wobble. Natural variety comes from the relationship between thoughts, not from alternating sentence lengths by formula.
Do not make text less usable or less accessible in the name of sounding less AI-written. Removing needed headings, lists, descriptive links, citations, caveats, or next steps is not a style improvement.
The recurring problem is regularity and mismatch, not any one feature. Use em dashes where they belong; do not reach for them as a default connective. If you keep using the same punctuation move in the same role, vary it rather than banning it. In casual internet prose, paragraph-after-paragraph em dashes are now a socially recognized AI cue, so prefer commas, colons, conjunctions, subordinate clauses, or full stops unless the dash clearly earns its keep. A full stop is not the automatic replacement; sometimes the fix is to make the relationship between the clauses clearer. For temporary compound modifiers, hyphenate before the noun and usually open after it; do not let the model turn every compound into a hyphenated unit.
Core Rules
1. Anchor to the actual context before drafting
Decide what the text is, who it is for, what register it uses, what answer or next action the reader needs, and, in replies, what thread, person, or community it is responding to. A reply that could be pasted into any thread on the same topic will often read generic even if the prose is clean. Keep register stable across the piece.
2. Fit the format to the medium
Format is part of register. Over-structuring casual writing makes it feel templated. Under-structuring technical writing makes it harder to use. Match the format to the medium instead of obeying a global ban on bullets, headers, or emphasis.
3. Prefer concrete specificity over polished generality
Each substantial paragraph should carry at least one concrete anchor.
What counts:
- a proper noun the reader could look up
- a specific number that is not only a date or version
- a direct quote
- a named decision, moment, or thread
- a checkable detail
- in criticism, reporting, and reviews: a user-facing or otherwise observed detail of what changed in practice
What does not count:
many, various, several, a lot ofin ways that mattered, meaningful changes, broad implicationsthe standard X arc, the usual pattern, as is often the case- vague intensifiers in place of claims:
essentially, fundamentally, ultimately
- milestone names, dates, titles, or feature labels standing alone with no material consequence attached
If the most concrete thing in a paragraph is a name and a date, the paragraph is still probably too generic.
4. Specificity must be earned
When writing about real entities, milestones, people, dates, quotes, events, public metrics, planned releases, or numbers, prefer fewer verified facts to many guessed ones.
Do not use specificity theater: invented milestone names, suspiciously exact claims, synthetic quotes, or decorative factuality added only to avoid sounding generic.
Be especially careful with hidden-mechanism claims: internal logic, unseen motives, back-end behavior, or claims about what a system is "really" doing under the hood. If the reader could not observe it and you cannot verify it, do not narrate it as fact.
Do not launder analysis through vague authority. Avoid experts say, observers note, research suggests, or critics argue unless you can name the source, describe what it actually supports, and keep your claim within that support.
Treat exact quotes, close paraphrases, public metrics, future claims, and causal claims as high-fragility facts. If you write that a person said something, a metric moved, a release is planned, or one change caused another, you need source support for that exact claim. If the support is weaker, narrow the claim: coincided with, appeared alongside, was followed by, or cut the relationship entirely.
If you cannot verify a claim, attribute it, soften it, or cut it.
5. Use plain words. Allow ordinary repetition. Prefer verbs.
Do not chase synonyms for basic words like problem, change, system, work, or people. Repeat the ordinary word when it is the right word. Prefer we changed it to the implementation of the change, latency dropped to a reduction in latency was observed, and applying the rule to the application of the rule. Prefer actions happening to people over abstractions being observed by systems.
6. Cohere through reference and sentence shape
Use pronouns and continued reference when the reader can easily track them. Do not restate the full frame in every paragraph. Treat signpost openers like Furthermore, Moreover, Additionally, Importantly, and Notably as things to justify, not default sentence starters.
Let closely related thoughts share a sentence when the relationship is tight. Use coordination for equal-weight thoughts (and, but, so), subordination for unequal ones (because, although, when, if, which, that), and colons or semicolons when the second clause explains, sharpens, or turns the first. A period should mark a real pause, shift, or emphasis, not merely the place where an adjacent thought happened to arrive.
Do not turn every idea into its own sentence for crispness. Short sentences are useful when the break creates emphasis, gives the reader time, or marks a real turn; they become false crispness when neighboring sentences are only separated because the prose is afraid to keep moving.
7. Do not perform
Avoid keynote cadence, mission-statement phrasing, applause-line endings, and ceremonial wrap-ups. Also avoid service-desk tone: no Great question, Absolutely, or similar canned praise unless the situation clearly calls for it; no I hope this helps, Feel free to reach out, or similar canned closers unless the situation clearly calls for it. Start where the answer starts. Stop where the answer stops.
8. Calibrate confidence, stance, and voice to genre
Be confident where evidence is strong. Be explicit where it is weak or interpretive. If the genre normally carries a visible writer (review, opinion, comment reply, personal post), let the writer appear. If the genre normally aims at neutrality (summary, documentation, news-style reporting), do not inject first person or attitude just to make the piece feel human. If the subject naturally invites a view, do not sand everything down to evenly polite neutrality. If the subject does not require a view, do not manufacture one. For public, technical, product, or instructional writing, keep language globally legible and inclusive.
9. Show concrete things before generalizing
Do not open with abstract diagnosis when the reader has nothing concrete to attach it to. This is not a ban on leading with the conclusion in web, docs, email, news, or task-oriented writing; if you lead with the conclusion, make it concrete enough to be useful. Usually the order should be:
- what happened
- where the pattern appeared
- what constraint mattered
- what failed or changed
- what that seems to mean
10. Watch regularity
LLM writing often becomes suspicious when its most visible feature is its own regularity.
Watch for repeated use of the same moves:
- parallel enumeration and reflexive three-part cadence inside sentences
- multiple sentences doing hidden list work even without bullets
- concession-plus-positive rhythm (
not X, but Y; may sound X, but Y)
- paragraph-closing type definitions (
the kind of X where Y)
- identical paragraph arcs
- one neat claim sentence at the top of every paragraph followed by orderly elaboration
- the same punctuation move in every paragraph
- the same controlling metaphor or contrast returning until it feels too tidy
- repeated thesis-like openings
- stacked mini-sentences for impact, especially when each sentence carries one adjacent thought that could have shared a sentence
Three-item parallel lists still count. The fix is not random variation; it is to break the repeated pattern where it starts to dominate.
One common over-correction is false crispness: splitting every clause into its own sentence to break regularity. For when to combine and when to keep the split, see rule 6.
11. Let the thought develop
Longer pieces should not feel pre-solved. If the prose moves in a perfectly efficient straight line from claim to conclusion, it can feel rushed. Let the thought develop through a concrete example, a noticed detail, a sentence that gathers related material, or a brief doubling-back when the material naturally allows it.
Development can happen inside a sentence, not only across paragraphs. A cumulative sentence can start with the main claim and then add the reason, qualification, or consequence that belongs with it. Do not split that movement into separate sentences unless the break itself does useful work.
12. Choose structure consciously for longer pieces
Default genre shapes are not wrong. They are only a problem when used by reflex.
For task pages, procedures, reference docs, and news briefs, the predictable structure is often the clearest one. Do not avoid it for novelty.
For retrospectives, criticism, feature writing, and other developmental pieces, the obvious defaults are often weak:
- starting state -> changes -> verdict
- one topic bucket per paragraph
- one paragraph per named milestone
Choose a through-line instead: one constraint that started biting, one shift in what people actually had to do, one mismatch between promise and reality.
Useful alternatives: thematic instead of chronological, reverse-chronological, perspective-led, counterfactual, opinion-first, single-example-led.
13. Do not turn a piece into catalog prose or system-tour prose
If a paragraph is mainly names, milestones, categories, feature nouns, or system labels, it is probably catalog prose.
If each paragraph can be summarized with a single label such as background, mechanism, impact, response, ending, the piece is probably system-tour prose.
Do not give one paragraph to each milestone or one paragraph to each topic bucket unless that mapping is the actual point. Pick one change and trace its consequence. Cross-wire the piece so paragraphs depend on each other instead of sitting like labeled boxes.
14. Revise by reading and cutting
Re-read as a first-time reader. Cut anything that is auditioning. Cut sentences whose only job is to announce the next sentence. Collapse paragraphs that restate each other. Replace the most generic clause in the piece with something specific or delete it. Most edits should make the text shorter, but do not confuse concision with chopping: combining two tightly related sentences can be the cleaner edit when it restores the relationship between the thoughts.
Required Checks
Run during revision (workflow Step 4). For pieces up to about 150 words or three short paragraphs, run checks 1-3, 5, 7, and 10. For longer pieces, run all checks. Do not output the audit unless asked.
- Register fit. Does the format, punctuation, formatting, and level of structure match the medium and the user's request? For web, docs, or UI text, did you preserve scannability and accessibility instead of flattening the piece for style reasons?
- Concrete-anchor audit. For each substantial paragraph, point to one concrete anchor. In criticism, reportage, reviews, and analysis, at least one paragraph in the whole piece should be built around a single concrete example or observed consequence rather than category summary. If you cannot point to that paragraph, add one.
- Fact discipline. Pick the three most fragile factual claims in the piece: dates, milestone names, quotes, close paraphrases, public metrics, future claims, causal trend claims. If you cannot vouch for them, attribute them, soften them, or cut them. If a citation or source is present, confirm it supports the exact claim rather than a nearby topic.
- Source-fit check. For factual writing, check every exact quote, close paraphrase, public metric, planned/future event, and causal claim. Do not keep
X caused Y, X drove Y, or X proved Y unless the source supports the relationship. Use weaker relationship language only when that weaker claim is still accurate.
- Regularity and sentence-continuity tripwire. Name the single most repeated visible pattern in the piece. If the same move appears 3 or more times, or dominates two consecutive paragraphs, rewrite at least one occurrence. Also scan for false crispness: two or more neighboring short sentences whose thoughts are tightly related but split apart. If a comma, conjunction, subordinate clause, colon, or semicolon would express the relationship more naturally, combine one pair. If the period creates useful emphasis or clarity, keep it.
- Repeated-frame check. If a central metaphor, contrast, or wording family appears throughout the piece, decide whether it is a useful motif or a too-neat scaffold. Keep it only where it adds force; vary or cut the rest.
- Stance and voice. If the genre expects a visible writer or evaluative stance, state the writer's view in one sentence to yourself. If you cannot, add stance where it does real work. If the genre expects neutrality, did you keep it neutral?
- Developed thought. For any piece longer than four paragraphs, identify one place where the prose pauses, doubles back, or notices a concrete detail off the main line. If the piece runs in a perfectly straight line from claim to conclusion, see whether one example or noticed detail would make it less pre-solved.
- Shape and spine. For any piece longer than three paragraphs, state the organizing principle in five words or fewer and the controlling claim in one sentence. If the shape is basically
starting state -> changes -> verdict, if paragraphs map one-to-one with named milestones, or if each paragraph is just one labeled topic bucket, restructure.
- Over-correction. Did you add fake-human moves (typos, slang, forced asides, random fragments, or artificial sentence-length targets) just to break a pattern?
These are tripwires, not goals. Use them to catch genericity, visible regularity, false specificity, and modular structure, not to manufacture variation for its own sake.
Optional Long-Form Diagnostics
Ten fallback heuristics for when the Required Checks above aren't enough for a longer piece — not targets to optimize for. See references/long-form-diagnostics.md.
Examples of Useful Corrections
- Generic -> specific. Avoid:
The change had broad implications across the team. Prefer: The change cut review time, but it also pushed more edge cases into the escalation queue.
- Puffery -> observable consequence. Avoid:
The project stands as a testament to the team's commitment to innovation. Prefer: The project reduced the weekly handoff from three meetings to one written checklist.
- Administrative detail -> material detail. Avoid:
The revision changed the process. Prefer: After the revision, decisions stopped being a silent queue in the background; someone had to choose what to slow down and what to push through.
- Specificity theater -> verified restraint. Avoid:
The February revision renamed the framework and rewrote intake handling. Prefer: Early revisions focused on intake edge cases and prioritization; if you cannot verify the milestone name or exact wording, leave it out.
- Hidden mechanism -> observable consequence. Avoid:
The internal logic finally understood what mattered. Prefer: After the change, obviously irrelevant outcomes stopped showing up in routine cases.
- Vague attribution -> supported claim. Avoid:
Experts say the redesign improved trust. Prefer: In the support queue, billing complaints fell after the pricing table stopped hiding plan limits.
- Causal overreach -> relationship restraint. Avoid:
The redesign drove trust higher. Prefer: After the redesign, refund questions fell in the support queue.
- Catalog prose -> argument prose. Avoid:
First came change A, then change B, then change C. Prefer: The important shift was not that the thing accumulated more pieces. It was that later changes finally introduced friction where the earlier version let people coast.
- System-tour prose -> cross-wired prose. Avoid: one paragraph for
background, one for process, one for impact, then a verdict. Prefer: trace one recurring constraint, show how it appears across the piece, and make the paragraphs depend on each other.
- Choppy -> connected. Avoid:
The term does real work. It names a pattern that was floating unnamed. Prefer: The term does real work: it names a pattern that was floating unnamed.
- False crispness -> carried relationship. Avoid:
The uncertainty is real. The confident register wrapping it is a default. Prefer: The uncertainty is real, but the confident register wrapping it is a default.
Audit Watchlist
These are not bans. They are quick places to scan when default LLM writing slips into formula.
Formula phrases and sentence moves to scrutinize
It's important to note that / It's worth noting thatWhen it comes to / In conclusionin today's fast-paced world / ever-evolving landscapedive deep into / embark on a journeynavigate used as a vague metaphorIt's not X, it's Y / Not because X, but because YWhat matters is... / The real issue is...This is not just..., it is...is a testament toserves as / stands as when is or has would be clearerplays a key role / plays a pivotal rolereflects broader, symbolizes, showcases, highlights, or underscores when attached to generic significance rather than evidence- vague source laundering:
experts say, observers note, research suggests, critics argue, many believe
- unsupported causality:
drove, proved, showed that, made clear that, tracked with, led directly to
- paragraph-closing type definitions (
the kind of X where Y)
- persuasive three-part cadence or triadic rhythm used by reflex (
clearer, faster, cheaper)
- one-thought-per-sentence strings where adjacent claims should be coordinated or subordinated
- decorative emoji and checkmark bullets in prose contexts
- internal tracker IDs in reader-facing prose (
ADR-NNNN, bd-xxxx) — these belong in commits and docs/decisions/, not human docs; describe the change, not its ID
- generic-to-the-platform replies that reference nothing specific to the actual conversation
Great question / Absolutely / I hope this helps / Feel free to reach out
Jargon defaults to scrutinize
Use only when they are plainly the right words, not because the model fell into them:
delve into, tapestry, realm, leverage, harness, foster, empower, unlock, unveil, vibrant, crucial, pivotal, compelling, robust, seamless, holistic, multifaceted, paradigm-shifting, underscore, testament, valuable insights, rich, profound, enhance, showcase, boast / boasts, ever-changing, ever-evolving, ever-growing
The problem is repeated fallback diction, not the existence of any one word.
Formatting artifacts in plain text to scrutinize
- smart quotes and curly apostrophes
- single-character ellipses
- other visible copy-paste formatting artifacts that do not fit the medium
Compound-modifier hyphenation to scrutinize
For temporary compounds, hyphenate before the noun: a well-known author, a high-quality service, a long-term plan. After the noun, usually open the compound: The author is well known, The service is high quality, The plan is long term.
Watch for:
- predicative over-hyphenation:
is well-known, seems high-quality
-ly adverb compounds: highly-qualified, newly-designed- reflexive
ever- compounds: ever-changing, ever-evolving
- set phrases where the hyphen adds nothing:
high school, ice cream
Do not strip all hyphens. Keep them when they prevent ambiguity, when the compound precedes the noun, or when the term is conventionally hyphenated. The problem is the reflex, not the mark.
Red Flags / Anti-Rationalization
| Thought | Reality |
|---|
| "The rules will make my writing worse" | The rules target LLM defaults. If you are writing well already, they will not change much. |
| "I'll skip the checks -- the draft is good enough" | The checks catch regularity and genericity you cannot see while drafting. Run them. |
| "Technical docs don't need prose quality" | Technical docs are the most-read prose in most organizations. Bad docs cost more than bad blog posts. |
| "I'll just run the watchlist as a find-and-replace" | The watchlist targets repeated fallback, not individual words. robust is fine once; robust in every paragraph is the problem. |
| "Sounding human means adding typos and slang" | Sounding human means fitting the context. Fake humanity is worse than default LLM prose. |
| "I'll strip all structure to avoid looking AI-generated" | Removing needed headings and lists from technical writing is not a style improvement. |
| "This piece is too short for checks" | Short pieces still run checks 1-3, 5, 7, and 10. The bar is lower, not absent. |
Capture what you learned. At close, record durable, evidence-backed insights (still true next month, tied to a file, test, or command). Never record guesses, one-offs, or secrets (tokens, keys, PII — every memory is injected into all future sessions). Update in place (bd remember --key <key>) rather than adding a near-duplicate.
bd remember "<kind>: <durable, evidence-backed insight>"
Integration
Pairs with:
document-release -- fires post-ship to sync docs to code; write-documentation fires when writing or rewriting prose. Complementary, not overlapping.verification-before-completion -- prose checks are part of completion evidencebrainstorming -- design docs written during brainstorming benefit from these rules
Called by:
- Any workflow where the agent is producing human-facing text
- User requests to write, draft, or rewrite documentation
