Agent Skills — a practical toolkit for AI coding agents
Four battle-tested skills for Claude Code, Codex CLI, and any agent that
supports the SKILL.md convention. Each one solves a failure mode that
actually costs time and money in daily agent work: vague requirements,
shallow code review, token waste, and deck grunt-work.

Skills
| Skill | One-liner | Highlights |
|---|---|---|
| prompt-forge | Vague request → engineering-grade prompt | Original CRAFT framework (Context/Role/Action/Format/Tests), ≤3 batched questions, 5-second verification test |
| prism-review | One change, four independent review lenses | Correctness / Security / Performance / Contracts run in parallel, noise-filtered severity merge, auto deep-review on auth/payment/crypto paths |
| token-warden | Token budget guardian across 4 consumption layers | Reading / Tool-output / Conversation / Generation routed to cheapest adequate tool; hard anti-pattern rules; zero-dep token estimator included |
| slide-smith | Batch PowerPoint automation | Deck inventory audit, design-token standardization (incl. accessibility preset), safe copy-never-overwrite workflow |
Install
Requires: nothing for prompt-forge / prism-review / token-warden;
pip install python-pptx for slide-smith.
Claude Code
cp -r skills/<skill-name> ~/.claude/skills/
Codex CLI
cp -r skills/<skill-name> "$CODEX_HOME/skills/"
Any compatible agent — copy the skill folder into your agent's skills
directory, or paste SKILL.md into your system prompt.
Install all four at once:
git clone https://github.com/Lesile-Yin/agent-skills.git
cp -r agent-skills/skills/* ~/.claude/skills/
Design principles
- One failure mode per skill. No Swiss-army skills; composition happens in the workflow, not inside a skill.
- Verifiable over vague. Every skill defines an output contract and refuses unfalsifiable claims (no invented percentages, no padding).
- Safe by default. Read-only review passes, copy-never-overwrite file ops, explicit escalation announcements.
- Cheap to run. Skills must not cost more tokens than they save; each one says when to skip itself.
Repository layout
skills/
├── prompt-forge/ SKILL.md · references/craft-framework.md · examples/
├── prism-review/ SKILL.md · references/lenses.md · templates/report.md
├── token-warden/ SKILL.md · references/routing-rules.md · scripts/estimate_tokens.py
└── slide-smith/ SKILL.md · references/design-tokens.md · scripts/{inventory,standardize}.py
Contributing
Issues and PRs welcome. Ground rules: keep one failure mode per skill, keep the output contract section, and no telemetry/network calls in scripts.
License
MIT © Lesile-Yin
No comments yet
Be the first to share your take.