Full dev cycle. Planning tier for clarify and plan; worker tier after approval for autonomous execution, tiered lint, thread rotation, and feedback fixes. Start every change here.
日本語の概要は準備中です。原文の説明を表示しています。
Writing anything the fleet produces for a person to read: a README, a changelog, a commit message, a review note, a task file, a docblock, a user-facing string. Carries the WordPress Documentation Style Guide rules in `references/`. Use before drafting or rewriting prose, then run the pass that takes the machine out: `humanizer` for English, `humaniseur-fr` for French.
インストールする前に、エージェントに与えられる指示の中身を確認できます。
A person reads this, not an agent. The goal is prose that sounds like someone wrote it, because someone did.
The fleet's register is US English, conclusion first, no preamble. Everything below is what that means in practice.
The fleet's standard is the WordPress Documentation Style Guide, and its rules are the house style wherever a person reads the output. Each heading below links to the reference that holds that section of the guide in full, with its examples and its exceptions. The lines are the whole guide in one line each; open the reference when a line is not enough to decide.
<var> element and use uppercase characters with
underscore delimiters.references/word-list/<first letter>.md. Open one letter, never the whole
dictionary.Seven references sit beside this file. What you are writing decides which one you open — not how big it is.
| reference | open it when |
|---|---|
references/general-guidelines.md | the question is structure, tone, audience, inclusivity, or a claim |
references/language-and-grammar.md | articles, capitalization, clauses, tense, voice, pronouns, word choice |
references/punctuation.md | apostrophes, colons, commas, dashes, hyphens, periods, quotation marks, slashes |
references/formatting.md | dates, filenames, headings, lists, notices, numbers, procedures, tables, trademarks, units |
references/linking.md | cross-references, heading anchors, image links, link text |
references/developer-content.md | code examples, inline code, command-line syntax, placeholders, UI elements |
references/word-list/<letter>.md | a term's spelling or usage — one file per first letter |
The guide governs documents and quoted text: em dashes for a break in the flow, straight apostrophes, straight double quotation marks. The humanizer lists govern prose habit: a dash used as a universal connector is a tell, and stacked punctuation is rationed. Both hold at once — documentation punctuation is house style, and the habit of reaching for the same mark is the tell.
Draft, then run the pass that matches the language: humanizer for an English
text, humaniseur-fr for a French one. Their lists are theirs and are not
restated here. Name the pass in the summary of anything you write about the text
— reviewer and qa record whether it ran, on any change that adds or rewrites
a sentence or more of reader-visible prose.
Could it be shorter? Most sentences can. The first draft explains; the second states. Cut the runway — "in order to", "it should be noted that", "it is important to", "the fact that".
Does it say anything? A sentence that survives deletion without loss was padding. The commonest padding is a restatement of the sentence before it, introduced by "in other words" or "essentially".
Would anyone say this? People do not "leverage functionality to facilitate outcomes". They use a tool to do a thing. If the sentence cannot be said aloud without embarrassment, it is written for a rubric rather than a reader.
The tells, in rough order of how often they appear here:
Lead with the conclusion. The reader decides in the first two lines whether to keep going. Say the thing, then explain it.
Be specific. "It broke" is not a bug report. "The second call returned an empty string when the cache was cold" is. Names, numbers, file paths.
Vary the length. A short sentence after two long ones lands. That is rhythm, and it is the difference between prose and a specification.
Say "you" and "we". The passive voice hides who did what: "the file was modified" is not a fact anyone can act on. "The merge rewrites it" is.
Own the uncertainty. "I think the second is faster" beats "the second may potentially offer improved performance". One is a claim, the other is a hedge wearing a claim's clothes.
| output | reader |
|---|---|
README.md, generated block and the project's own parts | a developer deciding whether to use this |
| commit messages | whoever runs git log in a year, probably you |
| pull-request prose | a reviewer deciding whether to read the diff |
changelog and readme.txt | a user deciding whether to update |
| review and QA findings | the person who has to fix them |
| task files | the next agent, who has only this to go on |
| code comments and docblocks | the next reader of the line above them |
| user-facing strings | someone reading the screen, in their own language |
Not the harness's own internal docs: DESIGN.md, the ADRs and the
migrations notes are written for whoever maintains the harness, and prose that
argues a decision is doing a different job from prose that reports one.
Read it aloud. If you would not say it that way to someone who asked, rewrite it until you would.
Distilled in part from mattpocock/skills/writing-for-agents, whose levers —
pruning, no-ops, leading words, negation — apply to prose as much as to
documents. See .agents/docs/DISTILLATION.md.
The humanizer-named skills on skills.sh that wrap a paid API
(humanizerai.com and similar) were rejected: they send the text off the
machine, which contradicts the fleet's first rule — everything runs locally. The
two that hold no API were distilled instead, as humanizer for English and
humaniseur-fr for French.
The rules above are the WordPress Documentation Style Guide, cut from
mirrors/wordpress-documentation-style-guide-consolidated.md by
mirrors/tools/build_style_guide_references.py and shipped in references/.
Every page in them keeps the live URL it came from on its **Source:** line.
This skill is still the one to reach for first. humanizer reads a draft and
takes the machine out of it; this one is how to write the sentence that does not
have the machine in it to begin with.
まだレビューはありません。使ってみた感想をお寄せください。
概要と使いどころ
Full dev cycle. Planning tier for clarify and plan; worker tier after approval for autonomous execution, tiered lint, thread rotation, and feedback fixes. Start every change here.
日本語の概要は準備中です。原文の説明を表示しています。
Writing documents an agent consumes. Use when creating or editing a skill, an AGENTS.md, a README section, a docblock, or any file reached by a pointer.
日本語の概要は準備中です。原文の説明を表示しています。
This repo is indexed by graft/. For ANY task here, whether understanding how something works, finding where code lives, tracing what calls a symbol or what a change breaks, or scoping an edit, get your context from the graph before grepping or reading source files. Every command runs through `bin/harness graft`.
日本語の概要は準備中です。原文の説明を表示しています。
Interviewing the owner to a shared understanding before any work starts. Use when a plan, a decision or an idea needs stress-testing, or when the owner asks to be grilled, interviewed or questioned about one.
日本語の概要は準備中です。原文の説明を表示しています。
Removing AI tells from French prose without lowering its register. Use when rewriting or reviewing French text — a reply to the owner, a `.po` target, a French post — that reads like a machine wrote it.
日本語の概要は準備中です。原文の説明を表示しています。
Editing English prose that reads as machine-written. Use when reviewing or rewriting a draft for AI tells — a README, a changelog entry, a comment, a user-facing string, a review. French text is `humaniseur-fr`.
日本語の概要は準備中です。原文の説明を表示しています。