Coleção Pessoal de Skills Claude Code
larendf/my-claude-skills
Gerencie e instale skills Claude Code com symlinks versionados
Ver no GitHubUm repositório versionado que centraliza skills próprias e de código aberto para Claude Code, instaláveis via script único. Cada skill é uma pasta com SKILL.md que ensina ao Claude uma capacidade focada (como escrever ADRs ou construir componentes), ativadas automaticamente por detecção de contexto ou invocação explícita.
Instalação
npx convez-ai larendf/my-claude-skills1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Organizar e reutilizar prompts especializados entre sessões de Claude Code
- Ativar automaticamente skills relevantes baseado em detecção de contexto
- Manter um catálogo versionado de padrões e receitas de desenvolvimento
Veredito

Projeto bem estruturado para quem trabalha frequentemente com Claude Code e quer evitar repetir prompts. O sistema de symlinks e descrição frontmatter é elegante, mas requer disciplina para manter as skills atualizadas e bem documentadas.
README
A tradução para o português deste README ainda está na fila.
my-claude-skills
My personal, combined collection of Claude Code skills — the ones I authored plus vendored open-source skills I rely on, all in one version-controlled place and installable with a single command.
A skill is a folder with a SKILL.md that teaches Claude Code a focused
capability (e.g. "write an Architecture Decision Record" or "build an AEM
component the Tourism Australia way"). Once installed, Claude pulls the right
skill into context automatically when your task matches it.
Quick start
git clone <this-repo> my-claude-skills
cd my-claude-skills
./install.sh # symlink skills, agents and commands into ~/.claude/
Then, in any Claude Code session:
- Just describe your task — Claude auto-activates a matching skill.
"Help me write an ADR for switching to Postgres" → the
adr-writerskill loads. - Or invoke one explicitly by name: type
/adr-writerin the prompt. - Not sure what's available? Type
/to browse, or skim the catalog below.
Skills are symlinked, not copied — edit a
SKILL.mdhere and the change is live in every session immediately. No reinstall needed.
Verify it worked
ls -la ~/.claude/skills # each entry should be a symlink → this repo
./install.sh --dry-run # re-run anytime to see what's linked vs missing
How skills work
- Discovery is automatic. Each
SKILL.mdhas adescriptionin its frontmatter. Claude reads those descriptions and loads a skill only when your request matches — so having 25 skills installed costs nothing until one is relevant. - Write good descriptions = better activation. The
descriptionis the trigger. The clearest ones say what the skill does and when to use it (most descriptions below end with a "Use when…" clause). If a skill never fires when you expect it to, tighten its description first. - Explicit beats implicit. When you want a specific skill, name it with
/<skill-name>rather than hoping the description matches.
Skills catalog
Grouped by category. The When to reach for it column is the short version of each skill's trigger.
aem — Tourism Australia AEM platform (mine)
| Skill | When to reach for it |
|---|---|
aem-build-deploy | Building/deploying any module, or troubleshooting a build (build.sh, Maven profiles, local localhost:4502). |
aem-component | Creating or editing an AEM component, dialog, or Sling Model across the four sites. |
aem-osgi-service | Writing a backend OSGi service, servlet, scheduler, workflow process, or event listener in a core/ module. |
aem-react-component | Adding or editing a React feature that mounts inside an AEM component (the tourismaustralia-react pipeline). |
ai-agents — building agent systems
| Skill | When to reach for it |
|---|---|
context-engineering | Writing or tuning commands, skills, and sub-agent prompts; managing the context window. |
multi-agent-patterns | A task exceeds one agent's context or splits into specialized subtasks. |
prompt-injection-defense | Vetting external CLAUDE.md/SKILL.md/MCP descriptions or sanitizing fetched/issue content before it enters context. |
rag-retrieval | Building or tuning a RAG pipeline (chunking, hybrid search, reranking, pgvector). |
devex — developer experience
| Skill | When to reach for it |
|---|---|
adr-writer | Recording a significant technical decision (Nygard-format ADR). |
ci-cd | Writing or improving a GitHub Actions pipeline. |
confluence-formatter | Turning a Markdown file or code into Confluence Cloud paste-ready, nicely styled Markdown (emoji, status circles, callouts, tables). |
playwright-e2e | Writing, structuring, or debugging Playwright E2E tests. |
structural-search | Precise, codebase-wide refactors by AST structure (needs the ast-grep/sg CLI). |
frontend
| Skill | When to reach for it |
|---|---|
react-server-components | Server-first React in Next.js 16 App Router; server vs client decisions; hydration bugs. |
shadcn-ui | Adding/customizing shadcn/ui components in a React project. |
tailwind-v4 | Styling with Tailwind v4 or migrating a v3 config. |
zustand-state | Adding or refactoring client state with Zustand 5.x. |
ui-component-review | Skeptical review of a built component: spec, conventions, gates, and a check that every design value traces to a source. Ends in APPROVE / REQUEST CHANGES. |
design-fidelity-check | Comparing a built component to its design reference at 375/1280/1440 with measured px/hex deltas. |
fundamentals
| Skill | When to reach for it |
|---|---|
accessibility | Building or auditing any web UI against WCAG 2.2. |
api-design | Designing or reviewing a REST/GraphQL API surface. |
database-patterns | Designing schemas or evolving a database with safe migrations. |
debugging | Tracking down a bug of unknown cause (leaks, races, deadlocks, crashes). |
security-patterns | Handling auth, untrusted input, or sensitive data (OWASP Top 10, secrets, PII). |
typescript-patterns | Modeling types, tightening strictness, or untangling complex generics. |
workflow — the learning loop
| Skill | When to reach for it |
|---|---|
record-mistake | Something went wrong: write a blameless entry whose Prevention line is a rule (docs/mistakes/). |
record-learning | You verified a non-obvious fact worth starting from next time (docs/learnings/). |
promote-rule | The same lesson keeps recurring: draft it as a checked-in rule, with a human approval gate. |
These generalize the equivalent skills in a project that already has them. Their names
differ on purpose (record-*, promote-rule, ui-component-review, design-fidelity-check)
so they never collide with a project's own log-mistake, propose-rule, review-component
or fidelity-check.
personal
| Skill | When to reach for it |
|---|---|
proposal-writer | Drafting client proposals, quotes, scopes of work, or engagement letters. |
task-prioritization | Ranking features or justifying roadmap order (RICE, WSJF, ICE, MoSCoW). |
Agents and commands
Besides skills, the repo carries user-level subagents (agents/*.md) and
slash commands (commands/*.md). install.sh links each file into
~/.claude/agents/ and ~/.claude/commands/, so they are available in every project.
| Agent | When to reach for it |
|---|---|
plan-roaster | A plan, spec or ADR looks finished and you want a cold, read-only second opinion: one weak point, failure modes, a score. |
design-critic | A UI was built or changed and you want severity-ranked, file:line findings on hierarchy, spacing, contrast, states and design-system use. Read-only, code only. |
security-reviewer | Auth, API endpoints, input handling, secrets or dependencies changed, or you want an OWASP-style pass. Can run npm audit and grep. |
| Command | What it does |
|---|---|
/component-pipeline <component or spec> | Plan check (plan-roaster) → build → parallel independent review (design-critic, security-reviewer) → verification gate run by a fresh agent (ui-component-review, design-fidelity-check) → at most 2 fix rounds, re-reviewed by the same reviewers. The builder cannot downgrade a finding. Leaves everything uncommitted. |
Keep project-specific agents (those that name a repo's paths, contracts or tools)
in that project's own .claude/agents/; only generic ones belong here.
prompting / sitecore
| Skill | When to reach for it |
|---|---|
promptify | Writing, fixing, or optimizing a prompt for a specific AI tool. |
sitecore-xmcloud | Sitecore XM Cloud / JSS work — auto-invoked in Sitecore projects. |
The aem/* skills are mine and encode Tourism Australia conventions (multi-site
AEM 6.5, Sling Models, OSGi R6, the React feature pipeline, build.sh).
Install reference
Each skill, agent and command is symlinked into your Claude directory so edits here are live everywhere. Use the script for your platform.
The Windows scripts (
install.ps1/uninstall.ps1) still handle skills only. Agent and command support is macOS / Linux for now.
macOS / Linux
./install.sh # symlink everything (skips entries already linked here)
./install.sh --dry-run # preview, change nothing
./install.sh --force # also replace existing links that point elsewhere (real files/folders are skipped)
./install.sh --replace-real # as --force, and back up + replace real files/folders too
./uninstall.sh # remove only the symlinks this repo created
If a name already exists, the script says whether it is a link to somewhere else
(and where) or a real file/folder, and skips it unless you pass --force. --force
replaces foreign links only. A real file/folder (for example a git clone you keep
updating) is left alone unless you also pass --replace-real, which moves it to
~/.claude/backups/my-claude-skills/<kind>/<name>.<timestamp> rather than deleting it.
Check with --force --dry-run first.
Custom locations: CLAUDE_SKILLS_DIR, CLAUDE_AGENTS_DIR, CLAUDE_COMMANDS_DIR
(backups: CLAUDE_BACKUP_DIR), e.g. CLAUDE_SKILLS_DIR=/path ./install.sh.
Windows (PowerShell)
.\install.ps1 # symlink all skills (skips ones already linked)
.\install.ps1 -DryRun # preview, change nothing
.\install.ps1 -Force # replace existing entries, even real directories
.\uninstall.ps1 # remove only the symlinks this repo created
Creating symlinks on Windows needs one of:
- Developer Mode — Settings → Privacy & security → For developers → turn on Developer Mode (no admin rights needed afterward), or
- an elevated PowerShell session (right-click → Run as Administrator).
If PowerShell blocks the script, allow it for the current session with
Set-ExecutionPolicy -Scope Process -ExecutionPolicy Bypass.
Install to a custom location with $env:CLAUDE_SKILLS_DIR = 'D:\skills' before
running the script.
Using WSL or Git Bash on Windows? Use the macOS / Linux
./install.shinstead.
Both platforms install the same skills — they only differ in how symlinks are
created. The target defaults to ~/.claude/skills (%USERPROFILE%\.claude\skills
on Windows).
Layout
skills/
<category>/
<skill-name>/
SKILL.md # required — frontmatter (name, description) + body
... # optional references/, scripts/, examples/, etc.
agents/
<agent-name>.md # subagent: frontmatter (name, description, tools, model) + prompt
commands/
<command-name>.md # slash command: /<command-name>
install.sh discovers skills by finding every SKILL.md, so nesting depth and
new categories work automatically — just drop a folder under skills/.
Add your own skill
-
Create
skills/<category>/<skill-name>/SKILL.md. -
Give it frontmatter —
name(kebab-case, matches the folder) and adescriptionthat says what it does and when to use it:--- name: my-skill description: One line on what this does. Use when <the trigger situation>. --- # My skill Body: the guidance, patterns, and examples Claude should follow. -
Run
./install.shto symlink the new skill (existing links are skipped). -
Start a fresh Claude Code session and confirm it activates — describe a task that matches the description, or invoke
/my-skilldirectly.
Keep SKILL.md focused; push long references, scripts, or examples into
sibling folders (references/, scripts/, examples/) so the main file stays lean.
Troubleshooting
| Symptom | Fix |
|---|---|
| Skill doesn't activate automatically | Sharpen its description (add a clear "Use when…" trigger), or invoke it explicitly with /<skill-name>. |
Changes to a SKILL.md aren't picked up | Start a new Claude Code session — skills are read at session start. |
install.sh skipped a skill | A real directory already exists at the target. Re-run with --force to replace it. |
| Symlink fails on Windows | Enable Developer Mode or run PowerShell as Administrator (see above). |
| Want to undo everything | ./uninstall.sh removes only the symlinks this repo created — your other skills are untouched. |
Attribution & license
MIT. The vendored skills are authored by Mikul Gohil and kept with their original attribution — see ATTRIBUTION.md.