Fullstack ML/AI Agent Skills
A Claude Code plugin — a catalog of production-grade, task-focused skills for
fullstack ML/AI engineering. Each skill packages instructions plus optional
scripts/assets so Claude can execute repeatable engineering workflows. The
repository doubles as a one-plugin marketplace, so it installs with the
standard Claude Code plugin flow.
Skills
dead-code-audit — Evidence-first dead-code + dependency hygiene audit for Python repos with reviewable docs/audit/ artifacts.fastapi-architect — Scaffold, review, or refactor FastAPI services using uv, a src/ layout, versioned routers, and optional singleton clients.llm-evals-toolkit — Design eval plans, datasets, metrics, judge prompts, and regression gates for RAG/LLM systems.minimal-tests-audit — Propose minimal, strictly-necessary tests with audit/progress files and an optional apply phase.mlops-blueprints — Produce right-sized MLOps blueprints for pipelines, deployment, monitoring, and governance.python-best-practices — Apply production-grade Python standards (structure, typing, docs, tests, uv+taskipy+ruff, CI gates, security, performance).repo-codebook-generator — Produce a repository codebook (structure, one-line file descriptions, full source) while respecting .gitignore.ralph-codex — Scaffold a Ralph-style autonomous agent loop into your repo, ready to be driven by OpenAI Codex CLI (or Amp).
Install (Claude Code plugin)
Add the marketplace, then install the plugin:
/plugin marketplace add Y4rd13/fullstack-ml-ai-agent-skills
/plugin install fullstack-ml-ai-agent-skills@y4rd13-skills
Skills load automatically and are namespaced under the plugin, e.g.
/fullstack-ml-ai-agent-skills:fastapi-architect. Claude also triggers a skill on
its own based on each skill's description. After installing (or editing skills
during development) run /reload-plugins to pick up changes.
Usage
Invoke a skill explicitly:
/fullstack-ml-ai-agent-skills:fastapi-architect audit this FastAPI project and propose an objective refactor plan
/fullstack-ml-ai-agent-skills:llm-evals-toolkit propose an eval plan for my RAG pipeline with metrics and test cases
…or just describe the task in natural language ("audit this repo for dead code",
"scaffold a FastAPI service with a src/ layout") and Claude will select the
matching skill.
Repository layout
.claude-plugin/
├─ plugin.json # plugin manifest (name, version, metadata)
└─ marketplace.json # one-plugin marketplace catalog (source: "./")
skills/<skill-name>/
├─ SKILL.md # required (YAML frontmatter + instructions)
├─ scripts/ # optional helper scripts
├─ assets/ # optional templates
├─ references/ # optional reference docs
└─ knowledge/ # optional background
Bundled scripts and assets are referenced as
${CLAUDE_PLUGIN_ROOT}/skills/<skill-name>/…. Claude Code substitutes
${CLAUDE_PLUGIN_ROOT} with the plugin's absolute install path, so skills work
from any project directory regardless of where the plugin is installed.
Local development & testing
- Load without installing:
claude --plugin-dir .
- Validate manifests + skills:
claude plugin validate .
- Reload after edits in a session:
/reload-plugins
Requirements
- Claude Code (latest version, with the
/plugin command)
- Some skills run helper scripts that require Python >= 3.12 and uv
Development (lint/format)
- Lint (check):
uv run task lint
- Lint (fix):
uv run task lint_fix
- Format:
uv run task format
Contributing
Small, focused PRs are welcome. Keep each skill self-contained under
skills/<skill-name>/ with a SKILL.md. Reference bundled files via
${CLAUDE_PLUGIN_ROOT}/skills/<skill>/… (never ~/.codex, <SKILL_DIR>, or
{baseDir}). Run claude plugin validate . before opening a PR.
License
See LICENSE.
