agent-doc-discipline — скилл для Claude Code: обзор и установка
Скилл agent-doc-discipline задаёт правила оформления инструкций, спецификаций и файлов навыков для ИИ-агентов. Пригодится при составлении CLAUDE.md, чтобы агент понимал задачу с первого раза без истории чата.
MIT 196 установок ★ 40 997 обновлён 2026-10-24
npx skills add Yeachan-Heo/oh-my-claudecode --skill agent-doc-discipline -a claude-codeЧто делает agent-doc-discipline
Скилл вводит строгую дисциплину оформления документов, по которым действует агент: файлов CLAUDE.md, GLOSSARY.md, стандартов в docs/standards/, дизайн-систем и навыков в .omc/skills/. Главный критерий прост. Новая сессия агента обязана выполнить задачу, опираясь только на текст, без ссылок на прошлые диалоги и контекст в памяти.
Внутри собраны практические требования к структуре. Шаги всегда ставятся в начало текста, а справочные детали сдвинуты вниз. Процедуры от трёх шагов обязательно нумеруются. Каждый шаг завершается чётким критерием готовности, иначе модель завершит работу раньше времени. Любое правило обязано содержать проверяемое условие и причину («why»), а списки должны умещаться в один взгляд.
Запрещено копировать в документацию то, что агент сам узнаёт из окружения: вывод --help, структуру папок или скрипты package.json. Для чисто человеческих текстов вроде описаний бизнеса в docs/business/ или историй выбора в ADR правила скилла не применяются.
Какие задачи решает
- Формулировать инструкции для агентов без опоры на контекст прошлых сессий
- Добавлять к каждому правилу проверяемое условие и обоснование
- Выносить первые шаги действий на начальный экран перед справочной частью
- Очищать документацию от дублирования вывода
--helpи структуры папок - Нумеровать процедуры от трёх шагов с критериями завершения
Примеры запросов агенту
- Перепиши наш
CLAUDE.mdпо правиламagent-doc-discipline, чтобы агент работал без контекста прошлых диалогов. - Отредактируй спецификацию в
docs/standards/: пронумеруй длинные процедуры и добавь критерии готовности. - Проверь описание навыка в
.omc/skills/и удали дублирование информации изpackage.json.
Как установить agent-doc-discipline
1Через Skills CLI (Vercel)
npx skills add Yeachan-Heo/oh-my-claudecode --skill agent-doc-discipline -a claude-code2Вручную: положить файлы
~/.claude/skills/agent-doc-discipline/SKILL.mdПапка скилла целиком; для одного проекта — .claude/skills/agent-doc-discipline/
1Через Skills CLI (Vercel)
npx skills add Yeachan-Heo/oh-my-claudecode --skill agent-doc-discipline -a codex2Вручную: положить файлы
.agents/skills/agent-doc-discipline/SKILL.mdДля всех проектов — ~/.agents/skills/
1Через Skills CLI (Vercel)
npx skills add Yeachan-Heo/oh-my-claudecode --skill agent-doc-discipline -a cursor2Вручную: положить файлы
.cursor/skills/agent-doc-discipline/SKILL.mdCursor читает и .agents/skills/, и .claude/skills/
1GitHub CLI
gh skill install Yeachan-Heo/oh-my-claudecode skills/agent-doc-discipline2Вручную: положить файлы
.github/skills/agent-doc-discipline/SKILL.md1Через Skills CLI (Vercel)
npx skills add Yeachan-Heo/oh-my-claudecode --skill agent-doc-discipline -a gemini-cli2Командой платформы
gemini skills install https://github.com/Yeachan-Heo/oh-my-claudecode --path skills/agent-doc-disciplineВручную: положить файлы
.devin/skills/agent-doc-discipline/SKILL.mdСтарый путь .windsurf/skills/ тоже читается
Через Skills CLI (Vercel)
npx skills add Yeachan-Heo/oh-my-claudecode --skill agent-doc-discipline -a opencodeClaude.ai
Упакуйте папку в zip и загрузите в Settings → Capabilities → Skills
По шагам, с командами для каждой платформы: Как установить скилл в Claude Code: пять способов
Вопросы об agent-doc-discipline
Как установить agent-doc-discipline в Claude Code?
Выполните команду npx skills add Yeachan-Heo/oh-my-claudecode --skill agent-doc-discipline -a claude-code. Файл навыка сохранится по пути ~/.claude/skills/agent-doc-discipline/SKILL.md.
Что делает скилл agent-doc-discipline?
Он вводит правила написания инструкций для ИИ-агентов в файлах CLAUDE.md, тикетах и спецификациях. Документы организуются так, чтобы агент из новой сессии мог действовать сразу по тексту.
Нужно ли применять этот скилл к документации для людей?
Нет, правила не распространяются на материалы исключительно для разработчиков, включая вводные руководства в docs/business/ и истории принятия архитектурных решений ADR.
Исходный текст
SKILL.md · MIT
--- name: agent-doc-discipline description: "Writing-time discipline for documents agents consume (the five surfaces, specs, tickets, .omc/skills/) — every rule checkable and carrying a why, steps before reference, one meaning in one home, no restating what the environment already says. Mandatory at drydock seed generation and the launch C5 sediment pass; opt-in for any other agent-facing doc edit. The companion of minimal-code-discipline: that one disciplines code, this one disciplines papers." level: 3 --- # Agent Doc Discipline Use this skill to apply a writing-time discipline while creating or editing any document an agent consumes to act: the five surfaces (`CLAUDE.md`, `GLOSSARY.md`, `docs/standards/`, `design-system/`, `.omc/skills/`), specs, tickets, and skill files. The test a document must pass: **a fresh agent session can act on it by reading alone.** Mandatory when: drydock generates surface seeds, the launch C5 sediment pass writes a lesson into a slot. Opt-in for every other edit — but any edit to an agent-facing document should survive these rules. ## When Not to Use Documents written for humans only: `docs/business/` narrative, ADR decision stories, READMEs for human onboarding. They may still follow the rules where it costs nothing, but they are not held to this discipline. ## The Discipline **Write for a reader with no chat history.** The document is self-describing: nothing leans on "as discussed", a prior session, or knowledge that lives in someone's head. _Test:_ a brand-new agent session can act on the document by reading alone. **Every rule checkable and carrying a why.** No "keep it clean" — write the observable condition and the reason. _Test:_ an agent can decide pass/fail from the text alone, and the why is stated next to the rule. **One meaning, one home.** Each rule or definition has a single authoritative place; duplication is drift waiting to happen. _Test:_ changing a behavior is a one-place edit. **Never restate what the environment confesses.** Package scripts, directory layouts, `--help` output are lookups, not documents — a copied lookup goes stale. Document only what cannot be found by looking: the unwritten convention, the reason behind a choice, the gotcha. _Test:_ every line survives "can this be looked up elsewhere?". **Steps first, reference behind, rare material behind a pointer.** What to do, in order, at the top; consult-on-demand rules below; low-frequency material pushed behind a pointer whose wording names the branches that should trigger reaching it. _Test:_ the next action is findable within the first screen. **Every step ends on a completion criterion.** Done must be tellable from not-done — a vague bound invites premature completion. _Test:_ for each step, "how do I know this is finished?" has an answer in the text. **A procedure longer than two steps is numbered.** The numbers are stable references: a reader can name step 3, follow the order, and jump without re-reading. An unnumbered procedure of many steps makes the reader hold the order in their head. _Test:_ every procedure of three or more steps carries its numbers. **A list is capped at what a reader holds in one glance.** A list that outgrows the reader's grasp loses its point — nothing in it is findable. A longer enumeration becomes structure (a table, subsections) or splits. This governs the writing of documents; it never touches the machinery of a gate or a batched decision, which lives in the methodology, not the prose. _Test:_ every list in the document fits in one glance, and any longer enumeration has become structure. **Deferred and scheduled work states when.** A document that defers work to a later step, session, or release states that timing next to the work it defers — "when" is a fact the reader needs to act, and an unstated deferral reads as an unstarted one. The timing is stated in the prose only; it never rebinds a checkpoint the methodology pushes right deliberately. _Test:_ every deferral in the document carries its when. **Close on a next action a reader can start now.** A report, ticket, or session-close pointer ends not with a summary but with the one next step a reader can begin in under two minutes — the destination, the pointer, or the command. A close that only summarizes makes the reader re-derive what to do first. _Test:_ the last line answers "what do I do next?" without scrolling up. **Re-state the position each time the reader rejoins.** A document a returning reader touches mid-flight (a long spec, a run report, a map Notes section) opens the new material with where things stand — step, phase, or decision count — before adding anything new. _Test:_ a reader arriving cold can locate the current state in the first two lines of the newest section. **Prompt the positive.** State the target behavior; a prohibition is reserved for hard guardrails and is always paired with the positive target. _Test:_ every prohibition in the document names the thing to do instead. **Scrape barnacles on write.** When the document contains stale or redundant material, remove it in the same edit; when it does not, add only the required material and do not invent deletions. A sentence the model already obeys by default pays load for nothing — delete the whole sentence. _Test:_ every line re-read earns its place against "does this change behavior versus the default?", and any stale or redundant material found during the edit is gone. ## Verification Before reporting a document change done, confirm: - a fresh session could act on it without asking a human anything - every rule is checkable and states its why; no rule restates a lookup - meanings live in one place; pointers name their trigger branches - stale or redundant material found during the edit was removed, while valid unrelated content was preserved - closes end on a next action the reader can start now; rejoining readers find the current state stated before new material - procedures of three or more steps are numbered; lists fit in one glance; every deferral states its when
© Yeachan-Heo · MIT · Yeachan-Heo/oh-my-claudecode