I'm a digital marketer, not a developer. But I work with coding agents every day, and the biggest upgrade wasn't a better model. It was skills.
What is a skill?
A skill is a folder with one file, SKILL.md. The top of the file has a short description that the agent matches against your request. The rest holds the instructions, loaded only when there is a match. No plugin, no framework, no code to run.
Say "cut this clip to 30 seconds and burn in captions" and an ffmpeg skill loads itself. Say "why is this page not converting" and a conversion skill loads instead.
What's in the collection
208 skills across these areas:
- Video and audio (ffmpeg, Remotion, transcription)
- Design and frontend
- Marketing and SEO (SEO audit, copywriting, ads, CRO)
- Docs and files (docx, pdf, pptx, xlsx)
- Backend and infra (Stripe, Supabase, auth)
- Agent workflow (planning, debugging, test-driven development)
- Security and reverse engineering
- Automation
I collected, edited and wired them together. Many are adapted from public work, including Anthropic's official skills, obra/superpowers, Corey Haines' marketing skills and OWASP cheat sheets. Every source is credited in the repo.
The real problem: skills fight each other
With 200+ skills installed, the problem is not "no skill". It is three skills firing at once. My rules:
- Load a skill only when the request clearly matches its description, not on one shared keyword.
- Max 3 skills per task. If two match, take the more specific one.
- No match? Work without a skill. A forced skill is worse than none.
I also rewrote descriptions so they work as triggers, not marketing copy, and added some triggers in Hinglish, because that is how I talk to my agent.
How to use it
You never need all 208. Copy the five you will actually use, and the agent's context stays small.
git clone https://github.com/Rachit-Sharma08/coding-agent-skills.git
Then copy a skill folder into ~/.claude/skills/ for Claude Code, or ~/.agents/skills/ for OpenCode. The full steps are in SETUP.md.
Repo: https://github.com/Rachit-Sharma08/coding-agent-skills
If it saves you an afternoon, a star helps other people trust the collection. And tell me in the comments: which skill is missing?

Top comments (2)
The 'descriptions are triggers, not marketing copy' point is the one most collections miss — I saw the same effect when trimming a large skill set: once descriptions overlapped on a keyword like 'review', the wrong skill fired even when a better one existed. Your max-3 rule plus 'work without a skill when there's no clear match' is the right guardrail; a forced skill is worse than none. One practical add-on that helped me: keep a short 'when NOT to use this' line in each description, which cut misfires noticeably. For teams working in Chinese, a similar curated approach exists in laolaoshiren/claude-code-skills-zh (github.com/laolaoshiren/claude-cod...), which pairs Chinese-language review and doc skills like zh-code-reviewer with a 400+ curated set, so trigger wording matches how the team actually phrases requests. Curious: did Hinglish triggers measurably improve hit rate over English-only descriptions, or was it more about your own phrasing habits?
Thanks, the "when NOT to use this" tip is great. I'm adding it to my descriptions.
Honest answer: I haven't measured Hinglish vs English properly. What I can say is that I prompt almost entirely in Hinglish and the right skill loads (for example my SEO audit skill fires from a Hinglish prompt), but that's one person's phrasing, not a controlled test. I also added a global rule in AGENTS.md (clear match only, max 3 skills, state which skill and why before loading), and I think that helps more than the language does. I'll try a proper side-by-side test and share the result.