GitHub · AI Programming · Programmer

Eight Agent Skill Repos That Keep AI Coders in Line

By 狂师 · Nov 3, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Agent Skills are turning into the standard way to enforce quality on AI-generated code, and these eight repos show that the discipline already runs across every major agent tool. Western teams can adopt the same workflow without waiting for vendor-specific features or proprietary frameworks.

Summary

Agent Skills are folders containing a SKILL.md that tell an AI agent when and how to use a capability, and they've become a cross-tool standard for Claude Code, Codex, Cursor, and WorkBuddy. Eight repositories stand out for enforcing engineering discipline: Karpathy's four rules stop AI from over-coding, Matt Pocock's grill-me interrogates the user before any work, and Ponytail's seven-step ladder pushes for the simplest possible solution without compromising validation or security. Addy Osmani's 25-skill set maps a full development lifecycle with specs, planning, TDD, review, and shipping.

For understanding existing code, Understand-Anything builds a knowledge graph from a 200,000-line repo, and Taste-skill forces real image sources and art direction for frontend design. Obsidian's CEO wrote official skills for markdown, canvas, and database files. Installation is uniform: Claude Code plugins or the skills CLI via npx, with impeccable offering its own installer. The real value, the article argues, is turning a senior engineer's judgment into an executable process—a team's own experience becomes a SKILL.md.

Takeaways

— Karpathy's guidelines repo (220k stars) enforces four rules: think before coding, simplicity first, surgical changes, and goal-driven execution.

— Matt Pocock's skills repo (300k stars) includes a grill-me skill that interrogates the user's plan step by step and a grill-with-docs variant that maintains a GLOSSARY.md.

— Ponytail (170k stars) uses a seven-step ladder from 'can we skip this?' to 'write minimal code' and reports 53% less code and 45% fewer tokens in the author's own 39-task benchmark.

— Addy Osmani's agent-skills repo has 25 skills grouped by lifecycle stage, including interview-me, spec-driven-development, and doubt-driven-development.

— Taste-skill (100k stars) forces a six-step design process using real images from public-domain sources like Wikimedia Commons and the Met Museum.

— Impeccable (90k stars) adds 62 deterministic checking rules run locally without LLM or keys, catching clichés like Inter-font-everything and purple-blue gradients.

— Understand-Anything (88k stars) builds a knowledge graph of a codebase using tree-sitter for structure and an LLM for semantics, outputting a JSON file team members can view without LLM access.

— Obsidian-skills (50k stars), written by Obsidian CEO Steph Ango, includes six skills for markdown, canvas, and database files, with defuddle and knap usable outside Obsidian.

— All eight repos install via Claude Code plugins or the npx skills CLI; impeccable has its own installer.

— The article recommends starting with a single skill you write yourself based on your own senior engineer's judgment, rather than installing many pre-made ones.

Conclusions

The common thread across these repositories is not making AI faster but making it do less, addressing over-engineering and hallucinated assumptions rather than throughput.

Standardization on SKILL.md means a skill written once runs on Claude Code, Codex, Cursor, and WorkBuddy, lowering the cost of sharing engineering practice across teams.

The article's closing argument, that copying a senior engineer's judgment into a SKILL.md is more valuable than installing ten pre-made skills, points to a future where teams author their own constraints instead of relying on external ones.

The rapid star counts (up to 300k) suggest that the demand is less for more AI capability and more for guardrails that keep AI from doing harm.

Concepts & terms

Agent Skill

A folder containing a SKILL.md file that describes what the skill does, when to use it, and how to use it. The agent decides on its own whether to invoke it based on the task.

SKILL.md

A markdown file that defines a skill's purpose, conditions for use, and workflow steps. It became an open standard adopted by Claude Code, Codex, Cursor, and WorkBuddy.

skills CLI

A command-line tool (`npx skills add`) that installs skills into a project or globally. It supports 70+ agents and is the universal installation method across the article's repos.

Chesterton's fence

A principle from the agent-skills repo that says don't remove or simplify something until you understand why it exists. It's used in code-simplification skills to preserve behavior while reducing code.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗