convez.ai
SkillDevOps & InfraestruturaAvançado
ifalex

Claude Skills - Coleção de Skills para Agentes de IA

ifalex/claude-skills

Skills de IA com teste, documentação e saída determinística para DevOps

Ver no GitHub
InicialPluginScore · de 0 a 100

Repositório com skills prontas para uso com Claude Code e GitHub Copilot CLI, focando em geração de artefatos complexos como Helm charts production-grade. Inclui skills token-otimizadas e determinísticas com testes automatizados, padrões de segurança e validação com kube-linter.

Instalação

$npx convez-ai ifalex/claude-skills

1 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

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Gerar Helm charts production-ready com segurança hardened
  • Criar skills determinísticas e testáveis para agentes Claude
  • Automatizar geração de artefatos complexos com validação integrada

Veredito

Repositório técnico bem-estruturado de skills de alta qualidade. A série helm-chart é exemplar em bulletproofing, testes e determinismo, mas requer experiência em Helm e DevOps. Valor real para equipes que precisam orquestração Kubernetes robusta gerada por IA.

README

A tradução para o português deste README ainda está na fila.

claude-skills

A collection of personal Agent Skills for AI coding agents. Each skill is a self-contained SKILL.md (plus optional reference files) that works with Claude Code and GitHub Copilot CLI — same format, two hosts.

This repo is the source of truth. You edit skills here, then deploy them into your agent's skills directory with deploy.sh.

Skills

SkillWhat it does
helm-chart-creatorInterviews you, then generates a production-grade, self-contained Helm chart (Bitnami conventions inlined — no OCI dependency) with security hardening, multi-environment values (dev/uat/prod), a values.schema.json, and validates the output with kube-linter + kube-score. Built for non-experts: every question has a default, an example, and explain-on-demand. The interview is a blocking gate — it never assumes defaults silently. Also runs an improve mode that audits and hardens an existing chart without changing its structure, presenting a proposal list first. Output is deterministic (templates copied verbatim) with a test harness to verify it.
helm-chart-token-optimizedToken-lean, deterministic sibling of helm-chart-creator — same production-grade output, but generation runs through a bundled scaffold.sh so template bodies never pass through the model's context (~40K tokens/chart saved). Short tiered interview (essentials first, full list on demand), confirmed once; same blocking gate and improve mode. Output is byte-identical for identical answers (cp+sed, not model text). Every environment (dev/uat/prod) passes kube-score with 0 CRITICAL and kube-linter clean out of the box — identical security posture across envs (non-root UID ≥10000, NetworkPolicy on, always-pull, ephemeral-storage, replica-safe PDB); only scale/resources differ. Ships a no-model test suite (run-all.sh): determinism, placeholder/defaults parity, schema lint, and a kube-linter/kube-score quality gate.

Skill-creator evaluation — helm-chart-token-optimized

Scored against the superpowers:writing-skills (skill-creator) rubric.

Rating: 8 / 10

Dimension (skill-creator)ScoreNotes
CSO / description9Triggers-only, "Use when…", no workflow summary. Verified it no longer lets models shortcut the Step 0 gate. Keywords + helm-chart-creator disambiguation present.
Naming / discovery9helm-chart-token-optimized is descriptive and intent-revealing.
Structure9Overview w/ core principle, When/When-NOT, quick-ref tables, Common Mistakes, Reference Files index — all present and well-formed.
Bulletproofing10Red Flags STOP table, rationalization→reality columns, explicit loophole closure ("Generate now ≠ ask nothing", "low effort budget never removes the gate").
Testing (Iron Law)8Behavioral RED/GREEN gate test baselined on the right tier (Sonnet) and verified live. Docked because it was retrofitted after the skill, not test-first; weak-model smoke check documented but not yet run.
Reference / file org10Lazy-loading table is exemplary — none of the reference files load on the common path. Heavy templates/explanations correctly behind branches.
Token efficiency5The drag. SKILL.md is ~2,690 words (~5× the <500 target) and loads every invocation. Some is justified (the gate must be inline), but the "always interview" point is restated ~5× and the generation/quality sections duplicate scaffold.sh --help.
Anti-patterns9No narrative storytelling, no multi-language dilution, no code-in-flowcharts. One excellent path (scaffold) over many mediocre ones.

Path to 9: a body-trim pass (de-duplicate the "always interview" restatements and the scaffold-internals prose down to ~1,200–1,400 words) lifts token efficiency without touching the gate language that makes the skill bulletproof.

Install

Clone the repo, then deploy with the script. By default it publishes all skills to Claude Code; use --target to choose where they go.

git clone git@github.com:ifalex/claude-skills.git
cd claude-skills

./deploy.sh                       # all skills  -> Claude Code (~/.claude/skills)
./deploy.sh helm-chart-creator    # one skill   -> Claude Code
./deploy.sh --target copilot      # all skills  -> Copilot CLI (~/.copilot/skills)
./deploy.sh --target all          # all skills  -> both hosts
./deploy.sh --dry-run             # preview, change nothing

Re-run deploy.sh whenever you pull updates. It replaces the deployed copy with the repo version (your edits live in the repo, not in the agent's skills folder).

Manual install (no script)

A skill is just a directory containing SKILL.md. Copy it into the right place:

HostPersonal (all projects)Project-scoped
Claude Code~/.claude/skills/<name>/.claude/skills/<name>/
GitHub Copilot CLI~/.copilot/skills/<name>/.github/skills/<name>/
# Claude Code (personal)
cp -R skills/helm-chart-creator ~/.claude/skills/

# GitHub Copilot CLI (personal)
cp -R skills/helm-chart-creator ~/.copilot/skills/

Copilot CLI also reads ~/.claude/skills/ and ~/.agents/skills/ for personal skills, so a single Claude Code install is often picked up by both. The ~/.copilot/skills/ path above is the canonical Copilot location.

Usage

Both hosts use the skill's description for automatic discovery — just describe the task and the skill triggers. You can also call it explicitly:

  • Claude Code: Create a Helm chart for my service (auto), or invoke the helm-chart-creator skill directly.
  • Copilot CLI: Use the /helm-chart-creator skill to chart my service (a leading / plus the skill name selects it).
  • Improve an existing chart: point it at a directory that already contains a Chart.yaml (e.g. Harden my existing chart in ./charts/api). It audits the chart, shows a proposal list grouped by severity, asks only the questions it can't answer from the chart, and applies fixes in place without restructuring.

Two flavours of the Helm chart skill: helm-chart-creator (v1) writes the templates through the conversation; helm-chart-token-optimized (v2) hands a small answer file to a bundled scaffold.sh that renders on disk, so it uses far fewer tokens and is byte-deterministic — pick v2 for minimal token usage or CI-repeatable generation.

Verify what's installed:

ls ~/.claude/skills/      # Claude Code personal skills
ls ~/.copilot/skills/     # Copilot CLI personal skills

Repo layout

claude-skills/
├── README.md
├── deploy.sh                       # copy skills -> agent skills dir(s)
└── skills/
    ├── helm-chart-creator/         # v1: writes templates through the conversation
    │   ├── SKILL.md                # orchestration: interview gate, create/improve modes, quality gates
    │   ├── reference/              # loaded on demand by the agent
    │   │   ├── explanations.md     # question catalog + "what is X" explainers
    │   │   ├── improve-existing.md # improve-mode audit checklist + proposal format
    │   │   ├── values-and-schema.md
    │   │   ├── quality-gates.md
    │   │   └── templates/          # every chart file (CHARTNAME placeholder)
    │   └── tests/                  # platform-neutral determinism harness
    │       ├── compare-charts.sh   # diff two charts; PASS if ≥95% identical
    │       ├── run-determinism.sh  # run a scenario N times on any agent CLI, then compare
    │       └── scenario-a.*        # fixed interview answers (prompt + human-readable)
    └── helm-chart-token-optimized/ # v2: renders on disk via scaffold.sh (token-lean, deterministic)
        ├── SKILL.md                # tiered interview, --plan confirm, scaffold-based generation
        ├── scaffold.sh             # deterministic cp+sed generator (templates never enter context)
        ├── answers.example.env     # fill-in answer file (the only thing passed through context)
        ├── reference/              # explanations, improve-existing, values-and-schema, quality-gates, templates/
        ├── docs/further-proposals.md
        └── tests/                  # no-model guard suite (run-all.sh)
            ├── run-scaffold-determinism.sh # scaffold twice, byte-compare (+ render overlays)
            ├── check-placeholder-parity.sh # %%X%% in tmpl <-> `add X` in scaffold.sh
            ├── check-defaults-parity.sh    # answers.example.env keys <-> scaffold.sh vars
            ├── lint-all-features.sh         # all toggles on; helm lint vs values.schema.json
            ├── check-kube-quality.sh        # kube-linter clean + kube-score 0 CRITICAL (uat/prod)
            └── check-deployed-sync.sh       # deployed copy vs repo (drift guard)

Adding a new skill

  1. Create skills/<name>/SKILL.md (see the Agent Skills spec for frontmatter).
  2. Keep SKILL.md lean; put heavy reference material in sibling files it links to.
  3. ./deploy.sh <name> to publish, then test it in your agent.

License

MIT — see LICENSE if present, otherwise use freely.

Ver o original no GitHub

Bom para

Relacionados