Claude Skills
eprouveze/claude-skills
Skills multi-LLM para planejamento, avaliação e operações de código
Ver no GitHubColeção de skills práticos para Claude Code cobrindo workflows multi-LLM, avaliação de planos, qualidade de conteúdo e operações de domínio. São ferramentas reais usadas em produção, com suporte a benchmarking de modelos, revisão de código, geração delegada e gerenciamento de domínios via API.
Instalação
npx convez-ai eprouveze/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
- Benchmarking e comparação entre múltiplos modelos de IA em tarefas de código
- Planejamento de projeto com exploração paralela do codebase
- Revisão de código independente e pair programming com segundo modelo
- Limpeza de texto gerado por IA e pesquisa de palavras-chave para SEO
- Gerenciamento de domínios e operações DevOps via APIs
Veredito

Conjunto maduro e pragmático de skills que tira proveito real de workflows multi-LLM. As ferramentas são diretas e resolvem problemas concretos (benchmarking, planejamento, revisão), com instalação mínima para a maioria. Recomendado para quem já usa Claude Code e quer automatizar operações técnicas complexas.
README
A tradução para o português deste README ainda está na fila.
Claude Skills
A collection of practical Claude Code skills built around multi-LLM workflows, plan/coverage checking, content quality, and domain ops. MIT licensed.
These are skills the author uses daily, cleaned up for public consumption. They are not polished products — they are working tools with rough edges, shipped with a feedback loop so the rough edges get filed down over time.
New here?
A skill is a folder of plain-text instructions that Claude Code reads and follows.
You trigger one by typing its name with a slash — /council, /deslop, /plan. Some
skills also ship a small helper script.
To use any of these you need Claude Code installed. The prompt-only skills then work right away — no terminal, no git, no keys. A few call an extra tool, key, or model; the next table says which. New to the idea of skills? See the official Agent Skills guide.
Want to just try one? Pick a ✅ skill in the table below and follow the easy install — those need nothing beyond Claude Code.
What's inside
| Skill | What it does |
|---|---|
battle | Benchmark AI model combinations on the same coding task. Solo + pair + leaderboard. |
brief | Project-aware session briefing. Loads project-intel, decisions, deferred actions. |
codex-write | Delegate heavy code generation to Codex CLI while Claude orchestrates and reviews. |
council | Five-model strategic advisory with role-locked seats and a synthesis pass. |
crawl | Tiered web fetcher → clean markdown. Direct fetch (zero-setup) + optional Cloudflare. |
deslop | Strip 22 common AI-writing tells from drafts (em-dashes, gift-wrapped endings, etc.). |
evaluate-plan | Coverage check on an implementation plan against its source requirements. |
keyword-research | Google Trends interest + related queries for SEO and content prioritization. |
mode | Toggle single-LLM vs multi-LLM routing for skills that support delegation. |
model-scan | Scan provider APIs and docs to keep a project's model table current. |
namecheap | Domain management via Namecheap's XML API (check, register, DNS, transfer, renew). |
pair-session | AI pair programming: Claude builds, a second model advises. Three styles. |
plan | Research-first planning: parallel sub-agents explore the codebase, then a plan doc. |
salesforce-reports | Create/clone/run/delete Salesforce Reports via the Analytics REST API + sf CLI. |
second-opinion | Independent code review via Codex CLI. Review, challenge, and consult modes. |
update-machine | Safe parallel package sweep (brew/npm/pipx/uv) with accumulated upgrade-trap guards. |
What each skill needs
Most need nothing beyond Claude Code. A few call another model, a key, or a Node package.
| Skill | Claude Code alone? | For the full thing |
|---|---|---|
brief, deslop, evaluate-plan, mode, plan | ✅ | — |
crawl | ✅ basic | Node (npx tsx); Cloudflare keys for JS-heavy pages |
update-machine | ✅ | the package managers you already use (brew/npm/pipx/uv) |
battle | ◑ partial | Antigravity + Codex CLI to benchmark the full model field |
council | ◑ partial | access to the other models (their CLIs or API keys) |
pair-session | ◑ partial | a second-model CLI (Antigravity or Codex) |
codex-write, second-opinion | — | Codex CLI |
keyword-research | — | Node + the google-trends-api package |
model-scan | — | provider API keys, plus Node |
namecheap | — | a Namecheap API key |
salesforce-reports | — | the sf CLI and a Salesforce org |
✅ works as-is · ◑ works, but better with extra models · — needs the listed setup first
Installation
The easy way (no command line)
You don't need git, or even a terminal.
Option A — let Claude Code install it. Open Claude Code and ask:
Install the
councilskill from https://github.com/eprouveze/claude-skills
Claude Code can fetch the files and put them in the right place for you. Swap council for
whichever skill you want.
Option B — download the ZIP.
- Click the green Code button near the top of this page, then Download ZIP (direct link: main.zip).
- Unzip it.
- Copy the folder of the skill you want — for example
skills/council— into your skills folder:~/.claude/skills/for all projects, or.claude/skills/inside one project. - Restart Claude Code.
With git (for developers)
Global (all projects):
git clone https://github.com/eprouveze/claude-skills.git ~/.claude/skills-source
# Symlink or copy individual skills into ~/.claude/skills/
ln -s ~/.claude/skills-source/skills/council ~/.claude/skills/council
Per-project:
git clone https://github.com/eprouveze/claude-skills.git
cp -r claude-skills/skills/council .claude/skills/council
After installation, restart Claude Code so the skill catalog refreshes.
Skills with their own scripts include a --setup flow:
~/.claude/skills/namecheap/scripts/nc_api.sh --setup
~/.claude/skills/model-scan/scripts/model-scan.ts --setup # via `npx tsx`
~/.claude/skills/salesforce-reports/scripts/sfreport.sh setup # installs sf CLI if missing
npx tsx ~/.claude/skills/crawl/scripts/crawl.ts --setup # optional Cloudflare creds
Before you run
These are real tools, not toys — a few can spend money or change things on your systems:
namecheapcan register, transfer, and renew domains, which bills your account.salesforce-reportscan delete reports in your Salesforce org.update-machineupgrades packages installed on your machine.
Treat them like any code off the internet: skim the skill's SKILL.md and any script under
its scripts/ folder before running. Every script accepts --help, and credential setup
writes to ~/.config/claude-skills/<skill>.env — never into the skill source.
Per-skill quick start
- battle —
/battle Write a token-bucket rate limiter middleware. Runs every available model and pair. Seeskills/battle/SKILL.md. - brief —
/brief. Loads project context for the current working directory. Seeskills/brief/SKILL.md. - codex-write —
/codex-write Refactor the checkout route to support multi-currency. Delegates to Codex CLI. Seeskills/codex-write/SKILL.md. - council —
/council Should we migrate from REST to GraphQL?. Five-seat advisory with synthesis. Seeskills/council/SKILL.md. - crawl —
npx tsx skills/crawl/scripts/crawl.ts "https://example.com". Direct fetch works with no setup; run... --setup(or setCLOUDFLARE_*env vars) to enable the JS-rendering tier. Seeskills/crawl/SKILL.md. - deslop —
/deslop content/blog/my-post.mdx. Scans for 22 AI-tell patterns. Seeskills/deslop/SKILL.md. - evaluate-plan —
/evaluate-plan @prd.md. Coverage report. Seeskills/evaluate-plan/SKILL.md. - keyword-research —
npm i google-trends-api, thennpx tsx skills/keyword-research/scripts/keyword-research.ts --keywords "a, b, c". Seeskills/keyword-research/SKILL.md. - mode —
/mode,/mode multi,/mode stats. Seeskills/mode/SKILL.md. - model-scan —
/model-scan. Refreshes the model table. Seeskills/model-scan/SKILL.md. - namecheap —
/namecheap check example.com,example.net. Seeskills/namecheap/SKILL.md. - pair-session —
/pair-session build Refactor the auth module. Seeskills/pair-session/SKILL.md. - plan —
/plan Add multi-currency support to checkout. Parallel research, then a structured plan indocs/plans/. Seeskills/plan/SKILL.md. - salesforce-reports —
scripts/sfreport.sh setup --org myorg, then... list/... clone --from <id> --name "Copy"/... delete <id> --yes. Org-agnostic; optional GAM Global Company filter. Seeskills/salesforce-reports/SKILL.md. - second-opinion —
/second-opinion review. Seeskills/second-opinion/SKILL.md. - update-machine —
/update-machine. Surveys + safely upgrades brew/npm/pipx/uv, holding session-critical casks for confirmation. Seeskills/update-machine/SKILL.md.
Conventions
--helpand--setup— every shipped script accepts--help; scripts that need config accept--setup(writes to~/.config/claude-skills/<skill>.envwhere it applies).- SPDX headers — every script carries
# SPDX-License-Identifier: MIT. - Env vars over config files — credentials live in env vars, not in skill source.
--setupflows write them to~/.config/claude-skills/<skill>.envso the user does not have to remember the names. - Frontmatter — every
SKILL.mdincludesversion,last-updated, andlast-consolidated. Bumpversionon body changes; refreshlast-consolidatedafter a learnings consolidation pass. - Tone — explain-the-why over MUST/NEVER imperatives. Specific, factual, no hype.
How these skills self-improve
Each skill ships with a lightweight feedback loop. Adopt it or ignore it — the skills work either way.
The pattern, in three pieces:
learnings.mdsidecar per skill. An append-only log of observations and corrections, dated, one paragraph each. When a user corrects an output, that correction lands here.- Consolidation pass (weekly or when the log crosses ~100 bullets). Each entry gets
one fate: apply (merge into the SKILL.md body under Known gotchas, Anti-patterns,
or Validated patterns), capture (leave it in the log), or dismiss (delete).
Bump
last-consolidated:in frontmatter. - Golden cases for skills with crisp pass/fail criteria (
namecheap,model-scan,evaluate-plan,council,battle). When a golden case that previously passed starts failing, that is a signal that consolidation needs to happen sooner rather than later.
Triggers to consolidate: explicit user correction (2–3 on the same theme is the loudest
signal), a novel input category recurring, the underlying model version changing, or a
breaking change in an external CLI the skill calls — including vendor renames (e.g., the
Gemini CLI → Antigravity CLI (agy) transition in May 2026, applied across this repo in
[v0.1.1: agy migration]).
The pattern is intentionally minimal. It does not require a runtime, a database, or any
tooling beyond text files. Edit learnings.md by hand; consolidate by hand; trust the
SKILL.md body as the source of truth.
Related
- heartbeat — an autonomous work loop for a Claude Code session: it wakes itself, reads ground truth, advances one item per tick inside a hard envelope, and queues anything irreversible to a human. It's a separate repo rather than a skill here because it's a small runtime system (a tick-protocol skill plus a launcher, board template, durability check, and an off-machine liveness watchdog).
Contributing
Adding a skill:
- Create
skills/<name>/SKILL.mdwith frontmatter:name,description,allowed-tools,version,last-updated,last-consolidated. - If the skill calls an external CLI or API, vendor any helper scripts under
skills/<name>/scripts/. Add SPDX headers and a--helpflag at minimum. - Include sections for Known gotchas, Anti-patterns, Validated patterns, and a
Self-improvement section pointing at
learnings.md. - Create
skills/<name>/learnings.mdfrom the template (see any existing skill). - If the skill has clear pass/fail criteria, add
skills/<name>/golden/with aREADME.mddescribing the cases.
Editing an existing skill: bump version, update last-updated, log the rationale in
that skill's learnings.md if the change came from an observation.
License
MIT. See LICENSE. Copyright Emmanuel Prouveze.
Credits
The self-improvement pattern is adapted from prior work by others — MindStudio's
learnings.md pattern, ChristopherA's bootstrap-skill gist (the rule-bloat warning),
Reflexion (verbal self-reflection), DSPy/BootstrapFewShot + MIPROv2 (gated promotion +
reserved baseline), and GEPA (natural-language critique over examples). The synthesis
above is mine; the underlying ideas are not.