claude-skills
adelrioj/claude-skills
Collection of Claude Code skills for dev usage
Instalação
npx convez-ai adelrioj/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
README
A tradução para o português deste README ainda está na fila.
Claude Skills
A Claude Code plugin bundling skills for autonomous story execution, adversarial spec review, and workflow support.
Skills
/plan-to-dex — Run a hardened Superpowers plan through the dex orchestrator. Translates the plan into dex's checkbox-group plan.md (one task = one iteration), imports it, and runs dex apply → dex review autonomously with codex as the fixed backend. Requires the dex and codex CLIs.
/swarm-execute — Implement a feature as parallel user stories: Claude orchestrates via the Workflow (ultracode) tool, Codex workers write all code in isolated worktrees, architect + QA reviews gate every merge. Takes a plain-language request or a plan/spec file directly — no prd.json or tasks/ files. Requires the codex CLI.
/spec-review-codex — Adversarial review of design specs using OpenAI Codex as an independent reviewer. Sends the spec to Codex for rigorous review against a 10-category checklist, fixes CRITICAL and IMPORTANT findings, and loops until the spec passes with zero blocking issues (max 3 iterations). Designed to catch bugs, contradictions, ambiguities, and gaps that self-review misses due to author bias.
/spec-review-local — Same adversarial review loop, but the reviewer is a local model served by LMStudio via the pi CLI. Auto-detects whatever LLM is currently loaded in LMStudio — no hardcoded model. Works fully offline; the reviewer is restricted to read-only tools.
/handoff — Compact the current conversation into a handoff document so a fresh agent can pick up the work. Writes to the OS temp directory, includes a "suggested skills" section, references existing artifacts by path instead of duplicating them, and redacts secrets/PII.
/sprint-status-update — Generate a company-wide sprint recap for Slack from the Notion sprint board: categorizes deliveries, summarizes bug reports, and formats a scannable update. Use on Fridays or at sprint boundaries.
/orbstack-compatible — Make a Docker Compose project stop colliding on host ports across worktrees/projects by moving containerized services onto OrbStack routable domains (<service>.<project>.orb.local). Strips ports: from the base compose, adds an opt-in docker-compose.ports.yml for Windows/CI, rewrites local env URLs to domains, then verifies live. Requires OrbStack as the active Docker engine.
Install
Add the marketplace, then install the plugin:
claude plugin marketplace add adelrioj/claude-skills
claude plugin install claude-skills@claude-skills-marketplace
Restart Claude Code for the plugin to load.
Update
claude plugin marketplace update claude-skills-marketplace
claude plugin update claude-skills@claude-skills-marketplace
Uninstall
claude plugin uninstall claude-skills@claude-skills-marketplace
To also remove the marketplace:
claude plugin marketplace remove claude-skills-marketplace
Local Development
For developing the plugin locally (session-only):
git clone https://github.com/adelrioj/claude-skills.git
claude --plugin-dir ./claude-skills
Spec Review
Two interchangeable skills run an adversarial review loop over brainstorming design specs (any file matching docs/superpowers/specs/*-design.md):
| Skill | Reviewer | Requirements |
|---|---|---|
/spec-review-codex | OpenAI Codex (codex exec) | Codex CLI installed and authenticated |
/spec-review-local | Local model via pi + LMStudio | pi in PATH, LMStudio running at http://127.0.0.1:1234 with an LLM loaded |
What happens:
- The skill locates the spec (argument, or the most recent
*-design.mdindocs/superpowers/specs/) - The spec is sent to the reviewer with a 10-category review checklist
- The reviewer verifies the spec's file paths and code references against the actual codebase and returns findings (CRITICAL / IMPORTANT / ADVISORY / MINOR) with a PASS or NEEDS REVISION verdict
- Claude fixes CRITICAL and IMPORTANT findings, then re-sends for review
- Loop continues until the reviewer returns PASS (max 3 iterations)
The loop runs autonomously — no approval prompts between iterations. Findings files are preserved in /tmp/spec-review-findings-<timestamp>.md for audit.
Invocation:
/spec-review-codex docs/superpowers/specs/2026-04-10-my-feature-design.md
/spec-review-local docs/superpowers/specs/2026-04-10-my-feature-design.md
Or without arguments to auto-detect the most recent spec:
/spec-review-codex
/spec-review-local