Claude Skills for Documentation
Claude skills for docs address one persistent failure: documentation gets written once at launch and then left alone until it has fully diverged from the code. A skill records how the document is structured and where its facts come from, so it can be regenerated on every meaningful change instead of rewritten by hand twice a year.
The most durable case is the README and the API reference. Their content is derivable from the code: endpoints, parameters, call examples. Once that derivation is written down as a procedure, updating docs stops being a separate task someone has to remember and becomes a side effect of changing the code.
The second case is the changelog. A skill assembles it from commit history without retelling it verbatim: commits are written for developers, changelogs are read by users. The rule for translating one into the other is exactly what the skill contains.
The collection below covers READMEs, references, changelogs and project knowledge bases.
Skills in this collection
- Очистка и полировка пользовательских текстов — clean-user-facing-text is a Claude Code skill that audits and rewrites authorized natural-language text, removing suspicious invisible Unicode characters while preserving every fact, number, name, citation, and requirement. It is designed for use in Cursor when polishing articles, manuscripts, reports, documentation, emails, product copy, UI text, Markdown, or HTML prose. The workflow has two stages: a deterministic Unicode pass using bundled Python scripts (`inspect_text.py` and `clean_text.py`), and a prose rewrite that varies clause order, sentence rhythm, and connectors to replace formulaic phrasing with direct, natural wording. Code blocks, commands, paths, identifiers, formulas, verbatim quotations, and required disclosures are never modified. The skill does not support code-only tasks and explicitly does not claim detector evasion or proof of human authorship.
- Ведение архитектурных решений (ADR) — documentation-and-adrs is a Claude Code skill that records architectural decisions and captures the context future engineers and agents need to understand the codebase. It is designed for moments when choosing between competing technical approaches, changing a public API, shipping user-facing features, or preserving reasoning that would otherwise be lost over time. The core artifact is an Architecture Decision Record (ADR): documents stored in `docs/decisions/` with sequential numbering, covering Context, Decision, Alternatives Considered, and Consequences, following the lifecycle PROPOSED → ACCEPTED → SUPERSEDED. Before generating a new ADR, the skill inspects the repository for existing conventions — directory, file format, numbering scheme, and section headings — and matches them rather than imposing its own template. Inline documentation guidance is also included, emphasizing comments that explain *why* rather than restating what the code already shows.
- Разработка по спецификации до написания кода — spec-driven-development is a Claude Code skill that produces a structured specification before any code is written, turning vague requirements into a shared source of truth agreed upon by the engineer and stakeholder. The workflow runs through four gated phases — Specify, Plan, Tasks, Implement — each requiring explicit human approval before proceeding. When a single request bundles several independently testable capabilities, an optional Phase 0 Scope Check kicks in: Claude proposes a capability map with a module table, dependency direction, and build order, then recurses through the full four-phase cycle for each module. The resulting spec document covers objective and acceptance criteria, full executable build and test commands, project structure, code style illustrated with real snippets, testing strategy, and a three-tier boundary system (always do / ask first / never do). Teams use this skill when starting a new project or feature, working from incomplete requirements, or facing architectural decisions that touch multiple files or modules.
- Разработка по официальной документации — source-driven-development is a Claude Code skill that grounds every implementation decision in official documentation rather than potentially stale training data. It follows a four-step process: detecting the exact stack versions from dependency files (package.json, requirements.txt, go.mod, and others), fetching the specific documentation page from an authority hierarchy (react.dev, docs.djangoproject.com, MDN, web.dev, and similar), implementing code against the documented API signatures, and citing sources directly in code comments so users can verify every pattern. When official sources conflict with each other or with existing project code, the skill surfaces the discrepancy and presents options instead of silently picking one. It is built for developers where correctness matters — especially when writing boilerplate, handling forms, routing, state management, or authentication with any framework or library.
- Шаблон пользовательской slash-команды для Claude Code — example-command is a Claude Code skill that serves as a reference template for building custom slash commands using the `skills/<name>/SKILL.md` layout. It documents every supported frontmatter field — `name`, `description`, `argument-hint`, `allowed-tools`, and `model` — and shows how to handle required and optional arguments via the `$ARGUMENTS` variable. The skill is pre-approved to use the Read, Glob, Grep, and Bash tools. Functionally it is identical to the legacy `commands/example-command.md` format; only the file structure differs. Developers creating their own Claude Code plugins or slash commands can use this as a ready-made starting point instead of piecing together the syntax from documentation.
- Шаблон для создания скилла Claude Code — example-skill is a Claude Code skill that serves as a reference template for building new plugin skills, covering their structure, file layout, and frontmatter format. The core SKILL.md file documents required and optional directory contents — including README.md, references/, examples/, and scripts/ subdirectories — along with frontmatter fields (name, description, version, license) and guidelines for writing effective trigger descriptions. It activates when a user asks to demonstrate skill format, create a skill template, or discusses plugin development patterns. Developers looking to build their own Claude Code skills will find clear explanations of how skills differ from commands and agents, plus concrete examples of trigger phrases that reliably invoke a skill.
- Аудит и улучшение CLAUDE.md файлов — claude-md-improver is a Claude Code skill that audits and improves CLAUDE.md files across an entire repository so Claude Code receives accurate, up-to-date project context. It operates in five phases: discovery of all CLAUDE.md, .claude.local.md, and nested variants via Glob and Bash; quality assessment of each file against six weighted criteria — commands/workflows, architecture clarity, non-obvious patterns, conciseness, currency, and actionability — scored A through F; generation of a detailed per-file report with a breakdown table; targeted diff-format improvement proposals with explanations; and Edit-based updates applied only after explicit user approval. The skill is designed for developers who want to fix stale build commands, document monorepo package relationships, or capture environment setup that Claude keeps missing — essentially turning CLAUDE.md maintenance into a structured, reviewable process.
- Аналитика сессий Claude Code в HTML-отчёте — session-report is a Claude Code skill that analyzes session transcripts from ~/.claude/projects and produces a self-contained interactive HTML report covering token usage, cache efficiency, subagents, skills, and expensive prompts. The bundled analyze-sessions.mjs script collects data for a configurable window — 24h, 7d, 30d, or all time — and outputs a JSON with sections for overall, by_project, by_subagent_type, by_skill, cache_breaks, and top_prompts. The skill copies a bundled template (with pre-built JS/CSS for sorting, bars, and drill-downs), injects the JSON blob, then fills an anomalies block with 3–5 findings expressed as percentages of total tokens and an optimizations block with 1–4 actionable suggestions tied to specific projects or prompts. It's built for developers and teams who need visibility into Claude Code spending, want to spot cache-hit rates below 85%, or identify subagent calls averaging over 1M tokens per invocation.
- oh-my-claudecode — справочник агентов и оркестрации — omc-reference is a Claude Code skill that serves as the built-in oh-my-claudecode reference: a catalog of 19 agents with model routing (haiku/sonnet/opus), tool documentation, a skills registry, and a structured commit protocol. It auto-loads when delegating to agents, using OMC tools, orchestrating teams, or invoking skills — keeping CLAUDE.md lean while detailed reference stays available on demand. The skill covers every agent role (explore, analyst, planner, architect, debugger, executor, verifier, and more), state and notepad commands, project memory, LSP and AST tools, and a five-stage team pipeline. The skills registry includes both workflow skills (autopilot, ralph, ultrawork, ralplan, deep-interview, ai-slop-cleaner) and utility skills. Built for developers who need centralized orchestration documentation without bloating the session prompt.
- Управление контекстом проекта через документацию — context-driven-development is a Claude Code skill that creates and maintains project context artifacts — product.md, tech-stack.md, workflow.md, and tracks.md — inside a conductor/ directory. It can scaffold greenfield projects from scratch, extract structured context from existing codebases, validate artifact consistency before implementation begins, and keep documents in sync as the project evolves. Teams use it for onboarding, defining development workflows, managing tech stack decisions, and ensuring that every AI-assisted session starts from a shared, up-to-date understanding of the project rather than guesswork or stale assumptions.
- Глубокая документация кодовой базы для AI-агентов — deepinit is a Claude Code skill that performs deep codebase initialization by generating hierarchical AGENTS.md documentation across every project directory. It traverses the directory tree level by level — excluding node_modules, .git, dist, and similar folders — then analyzes each directory and writes a structured AGENTS.md covering key files, internal and external dependencies, code patterns, and special instructions for AI agents. Every file except the root includes a `<!-- Parent: ../AGENTS.md -->` tag, building a fully navigable hierarchy. On subsequent runs, auto-generated sections are updated while manual annotations below the `<!-- MANUAL -->` marker are preserved. The skill delegates work across specialized subagents (explore, architect, writer) and supports parallel processing of same-level directories. It suits teams that want their codebase to be immediately readable by AI tooling and easier to navigate for both agents and new developers.
- Персистентная база знаний между сессиями — wiki is a Claude Code skill that implements a persistent markdown knowledge base compounding across sessions, inspired by Karpathy's LLM Wiki concept. It exposes six core operations: `wiki_ingest` (process knowledge into one or more pages), `wiki_query` (keyword and tag search with LLM-synthesized answers and citations), `wiki_lint` (health checks for orphan pages, stale content, broken cross-references, and structural contradictions), plus `wiki_add`, `wiki_read`, and `wiki_delete`. Pages are stored as YAML-frontmatter markdown files under `.omc/wiki/`, organized into categories such as architecture, decision, pattern, debugging, and session-log. Cross-references use `[[page-name]]` wiki-link syntax, and at session end significant discoveries are automatically captured as session-log pages — all without vector embeddings, relying solely on keyword and tag matching.
- Система памяти для писателей-фикшн — writer-memory is a Claude Code skill that provides an agentic memory system for fiction writers, persisting character profiles, emotional arcs, relationships, scenes, and themes across Claude sessions. All data is stored in a single `.writer-memory/memory.json` file designed for git-friendly collaboration. The skill exposes over 15 commands — including `init`, `char add/update`, `rel add/update`, `validate`, `synopsis`, and `export` — covering everything from tracking a character's forbidden words (taboo) and speech formality level to checking whether a line of dialogue fits that character's established tone. The `synopsis` command generates an emotion-focused story summary structured around five narrative elements. Korean-language workflows receive first-class support, with field labels such as 감정궤도, 대사톤, and 말투 레벨, making this skill especially suited for authors writing in Korean storytelling traditions.
- Подготовка к сертификации CMMC Level 2 — achieving-cmmc-level-2-compliance is a Claude Code skill that guides defense-contractor organizations through CMMC Level 2 certification: scoping CUI and FCI boundaries, implementing all 110 NIST SP 800-171 Rev 2 security requirements across 14 families, computing an SPRS score using the DoD Assessment Methodology, and building documentation for a C3PAO third-party assessment. The skill covers full workflow — classifying assets into scoping categories (CUI Assets, Security Protection Assets, Contractor Risk Managed Assets), authoring a System Security Plan, and managing a compliant POA&M under 32 CFR Part 170 constraints, including the minimum score of 88 for Conditional status and the 180-day closure deadline. Designed for Defense Industrial Base teams operating under DFARS 252.204-7012/7019/7020/7021 contracts that store or process Controlled Unclassified Information.
- Написание bug bounty репортов — report-writing is a Claude Code skill that guides bug bounty hunters through writing platform-ready reports for HackerOne, Bugcrowd, Intigriti, and Immunefi using an impact-first, human-tone approach. It provides dedicated report templates for each platform, a title formula, an impact statement formula, a CVSS 3.1 scoring guide, downgrade counter arguments, and a pre-submit checklist. The core rule is strict: no speculative language like "could potentially" or "may allow" — every claim must be proved or dropped. Each session ends with a complete findings/<target>-<bug-class>/ folder on disk containing the report draft, submission-notes.md, and an evidence/ directory, so nothing is lost in terminal scrollback. Ideal for security researchers aiming to increase report acceptance rates across major bug bounty programs.
- Сводный Markdown-отчёт из диагностических архивов — dbs-report is a Claude Code skill that merges multiple diagnostic snapshots saved by dbs-save into a single, shareable Markdown report suitable for clients, partners, or external consultants. It reads all archive files from `sessions/{project}/`, sorts them chronologically, deduplicates content, and organizes the output into six sections: evolution of the user's stated problem, confirmed conclusions, rejected directions, open questions, recommended next steps, and a full archive index. The skill supports `/dbs-report`, `/dbs-report --since YYYY-MM-DD`, and `/dbs-report --slug <project>` commands. Every generated report is written to `reports/{project}/` with a unique timestamp and never overwrites an existing file. It is designed for consultants and business analysts who need to consolidate incremental diagnostic sessions into a traceable, structured deliverable.
- Многослойная база знаний для личных и бизнес-решений — dbs-decision is a Claude Code skill that builds a local decision knowledge base for tracking long-term decisions across business, relationships, health, career, learning, and investments. It organises information into four strict layers: facts (`01_事实/`), patterns (`02_规律/`), state snapshots (`03_定格/`), and open questions (`04_待解/`), each with its own rules for writing and editing. The skill operates in five modes — project initialisation, current-state update, decision filing, result backfill, and periodic snapshot — and tags every entry with source markers such as `[AI 推测]`, `[结果回填]`, and `[修正]` to keep facts, judgements, and outcomes clearly separated. A privacy mode can be enabled at init to replace real names with aliases across all files. It is designed for anyone managing complex, multi-stage processes who needs to revisit past decisions months later with full context intact.
- Атомарная структуризация базы знаний и контента — dbs-content-system is a Claude Code skill that turns accumulated drafts, articles, case studies, course scripts, and social posts into a structured, reusable content engineering system designed to keep growing over time. It breaks content down into atomic units of five types — QST, CON, OPI, CAS, and SOL — then builds topic maps, tracks relationships, deduplicates entries, and supports Obsidian graph visualization. The skill ships with 10 Node.js tools covering scaffold initialization, unit draft generation, relationship mapping, duplicate detection, and assembly of new topic drafts, plus 7 templates and a complete ruleset. It is designed for situations where a large body of material already exists (50+ files or 80,000+ words) and the goal is a reusable content system rather than editing a single piece.
- Локальная база знаний для AI-агентов — dbs-knowledge is a Claude Code skill that turns a local folder into a knowledge base that AI agents can reliably search, update, and maintain. It covers four core operations: build — create a `SOURCE_OF_TRUTH.md` navigation file along with agent entry rules (`AGENTS.md`, `CLAUDE.md`); store — place new materials in the right location while resolving naming conflicts and version relationships; use — locate source files and answer questions with traceable references; and audit — detect broken paths, version conflicts, scattered files, and duplicate content. No database, vector store, embedding service, or third-party upload is required. The skill suits anyone who wants to organize a local document archive, enable AI-driven file retrieval, or keep a working folder structured without dealing with technical configuration details.
- Написание баг-репортов для bug bounty платформ — report-writing is a Claude Code skill that structures bug bounty reports for HackerOne, Bugcrowd, Intigriti, and Immunefi using an impact-first, human-tone approach. It includes platform-specific report templates, a title formula, an impact statement formula, CVSS 3.1 scoring guidance, severity downgrade counters, and a pre-submit checklist. The skill enforces a strict no-theoretical-language rule — every claim must be proven, never hedged with "could potentially." Each session persists the report draft, submission notes, and evidence under a findings/ folder on disk. Designed for bug hunters who have already validated a finding and need to write a clear, triager-ready submission.
- Умная база знаний из наблюдений ИИ — knowledge-agent is a Claude Code skill that builds and queries AI-powered knowledge bases from claude-mem observations. It follows a three-step workflow: `build_corpus` assembles a filtered corpus from observation history — filterable by project, type, concepts, file paths, semantic query, or date range, up to 500 entries — then `prime_corpus` loads it into an AI session, and `query_corpus` enables conversational Q&A with context preserved across follow-up questions. Corpora act as focused custom "brains": think "everything about hooks", "all decisions from last month", or "all bugfixes for the worker service". Commands `rebuild_corpus`, `reprime_corpus`, and `list_corpora` handle keeping knowledge current and managing session state. The skill suits developers who want to analyze past work patterns, recover context for specific modules, or compile expertise on a particular area of a codebase.
- Поиск по памяти прошлых сессий — mem-search is a Claude Code skill that searches the persistent cross-session memory database of claude-mem to retrieve work and context from previous sessions. It is designed for questions like "did we already fix this?", "how did we solve X last time?", or "what happened last week?" — cases where the current conversation has no answer but past sessions do. The skill enforces a three-layer workflow: `search` returns a lightweight index with IDs, timestamps, and titles (~50–100 tokens per result); `timeline` provides chronological context around an anchor observation; and `get_observations` batch-fetches full details only for the filtered IDs (~500–1000 tokens each), cutting token usage by up to 10x compared to fetching everything upfront. Filters include project name, observation type (bugfix, feature, decision, discovery, change), date range, and sort order. Developers maintaining long-running projects benefit most, especially when needing to avoid re-solving already-solved problems across multiple Claude Code sessions.
- Нарративный отчёт об истории проекта — timeline-report is a Claude Code skill that generates a comprehensive narrative report titled "Journey Into [Project]" by analyzing a project's entire development history stored in claude-mem's persistent memory timeline. It fetches the full timeline via a curl call to the claude-mem worker API, retrieving all observations, session boundaries, and summaries in pre-formatted markdown, then deploys a subagent for deep analysis. The report covers required sections including project genesis, architectural evolution, key breakthroughs, work patterns, and token economics — the last queried directly from the SQLite database at `~/.claude-mem/claude-mem.db`. Before processing, the skill estimates token consumption (ranging from roughly 20K to 750K tokens depending on project size) and asks for user confirmation when the timeline exceeds 100K tokens. It also handles git worktree detection to correctly identify the parent project as the data source. Ideal for developers and teams who need to document decision history, prepare project retrospectives, or reconstruct the full story of a codebase's evolution.
- Серийный дайджест истории проекта по неделям — weekly-digests is a Claude Code skill that generates a serial, multi-chapter narrative digest of a project's complete claude-mem timeline by splitting it into per-ISO-week files and running one consecutive subagent per week. Each subagent receives the prior week's carry-forward block so the story stays coherent across chapters — the total chapter count equals the number of ISO weeks the timeline covers. The skill auto-detects the project via git worktree logic, fetches the full timeline, splits it with a Python script that runs a dual-pass sanity check on observation counts, and builds a weekly index README as a roadmap. It targets developers and technical leads who need a structured, week-by-week story of how a project evolved — as opposed to a single sweeping report like timeline-report.
FAQ
How do I keep documentation up to date automatically?
The practical route is not to generate the whole document but to record in the skill which sections are derived from code, and regenerate only those when the relevant files change. Sections that explain reasoning and decisions stay human-written — they cannot be derived from code at all.
How is this different from generators like Swagger?
A generator extracts structure: endpoints, types, parameters. A skill supplies what the code does not contain — why the endpoint exists, which scenario it serves, and what people typically get wrong when calling it. They complement each other rather than compete.
All Claude Code skills