The 100K-star agent-skills repo is really about one thing: written workflows

Your bot does the job perfectly on Tuesday and invents its own version on Wednesday. Not because the model got worse. Because the instructions lived in your head, or scattered across five prompts, and each run had to guess what you meant. Over one hundred thousand developers ran into the same problem with their coding agents, and a repo called addyosmani/agent-skills just crossed 100,000 stars on GitHub by fixing it the boring way: writing the workflow down as a skill file.

What the trending repo actually does

agent-skills is a collection of 25 "skills" for AI coding agents, shipped as markdown files plus slash commands (/spec, /plan, /build, /test, /review, /ship) that map to the development lifecycle. Each skill encodes how a senior engineer does the work: spec before code, small atomic tasks, tests as proof, review before merge. They install into 70+ agents (Claude Code, Cursor, Codex, Copilot, and more) with a single npx command, and they work because they are specific: steps, verification gates, and the mistakes to avoid, all written down where the agent reads them every time. It is MIT licensed and free. Nothing about it is magic. The whole idea is that a workflow you write once beats instructions you rewrite every session.

The same idea in MyClawBots

The outcome is not "get an AI that knows engineering." It is this: a recurring job you describe once, in a file your agent reads on every run, so run 40 follows the same steps as run 1. In MyClawBots that file is SKILL.md: frontmatter with a name and description, then the operating core (purpose, workflow, output contract, operating rules). I keep my recurring sweeps, the Threads comment patrol and the promo-replies sweep, in exactly this form. The prompt stops drifting because the workflow is no longer in my head.

Paste this into your agent, replacing the bracketed part with the workflow you already do by hand:

Turn this recurring workflow into a skill file for my MyClawBots agent. Output it as SKILL.md with a YAML frontmatter block containing name and description, then these sections: Purpose, Workflow, Output Contract, Operating Rules.

1. Name it in lowercase-with-dashes, describing the single job it does.
2. Write the description as one sentence: what the skill does, and when the agent should use it.
3. In Workflow, write the steps exactly as I do this task today: [describe your workflow here, in plain steps, naming the tools, accounts, and checks it uses].
4. In Output Contract, define what the agent hands back: the format, one line of report, and anything it must keep in a running list so the same work never repeats.
5. In Operating Rules, list the hard constraints: what it must never do, what is a real failure worth retrying versus one to report and hold, and the privacy rule (no personal information in any public output, ever).
6. Keep SKILL.md short. Anything useful only sometimes goes in a references/ folder, not in the core file.

What a skill file cannot do

Honest limits. A skill file adds no tools and no connections. It cannot give your agent abilities the connected accounts do not already have; it makes the agent consistent, not capable. The skill only works if the agent actually loads it when the trigger matches, so the description line matters more than most people think: vague trigger conditions mean the skill sits unused. And agent-skills is built for coding agents, not for no-code bots. Borrow the principle, not the file.

Copy this to start

Pick the one recurring task you re-explain to your agent most often. Paste the prompt above, fill in your workflow, and run the resulting skill file for a week. If runs 1 through 7 look the same, you got the idea right. Flat pricing while you iterate: Lite $19, Pro $39, Max $79, Team $59 per seat, with a 3-day trial that needs no credit card.