Describe one task you do over and over. Get a starter SKILL.md you finish in two minutes, then paste straight into Claude or ChatGPT. The first skill in your company's library.
A SKILL.md is a markdown file that teaches an AI model how a specific task gets done in your business: when it triggers, the steps your best person follows, what good looks like, and the rules it must never break. Claude loads it as a skill. ChatGPT takes the same content as Custom GPT instructions. One file per repeated task, and the collection is your skill library.
This is the easy part made visible. Access (connectors, data) gets an agent into the work. A skill teaches it how the work actually gets done. Your private, company-specific skills are the asset competitors cannot download.
Pick a repeated, judgment-heavy task. The kind where your best person consistently gets a better result than everyone else. Answer what you can. Blanks are fine, the tool marks them so you know exactly what to finish.
Add a task name to generate. The rest is optional.
This is a starter you finish, not a finished skill. The tool scaffolds your inputs into the correct Anthropic SKILL.md shape and adds best-practice guardrails by keyword. The judgment is still yours to sharpen.
Everything below is what turns this starter into a real, reusable skill. Items flagged ! are gaps the tool found in your inputs.
Claude Code: add it to your skills directory as SKILL.md in a folder named after the skill, and Claude loads it as a skill. In a Claude Project, paste it as a custom instruction or knowledge file.
ChatGPT: paste the body into a Custom GPT's instructions, or drop the file into a Project's files.
Paths and menu names follow each platform's current docs, so check theirs if a step has moved.
Then run the task. Watch where the model gets it wrong, and feed that back into the rules and examples. That feedback loop is the whole game.
This tool builds one starter from one description. The real work is mapping every repeated, judgment-heavy task across your business, interrogating the people who do them best, and packaging each as a skill your agents can actually run. That map, and the install, is the Works Audit.
The Audit is where the keyword guesses this tool makes get replaced with your actual method, your edge cases, your quality bar. The gap you can feel between this starter and a finished skill is exactly the gap the Audit closes.
It does not write the skill for you. There is no AI running here. It takes your inputs and structures them into a correct, paste-ready SKILL.md, then adds a small set of curated best-practice guardrails by matching keywords across what you enter (email, report, data, client work, and a universal stop-guard so it never invents missing facts).
Output quality scales with input quality. Vague inputs make a vague starter. Specific inputs make a strong one. The keyword enrichment is a small, curated set, so it is helpful, not exhaustive. Treat the result as a starter you finish, not an AI-generated skill.
That gap is the point. Seeing how much of a good skill is your judgment, not the template, is the clearest demonstration of why the work, not the tool, is the asset.
Email yourself the file plus the finish-it checklist. We send it once, and follow up only if you want a real Audit. No list-bombing.
A skill library is the collection of SKILL.md files that captures how your business actually works: one skill per repeated, judgment-heavy task, each encoding the method of the person who does it best. Access (connectors, data) gets an AI into your work. Skills teach it how the work gets done to your standard.
The library compounds. Every skill you add sharpens the next one, and because the files are plain markdown they move with you across Claude, ChatGPT, or whatever model wins next. Your private, company-specific skills are the one asset a competitor cannot download.
Mapping that library across a whole business, then installing it, is what the Works Audit and Embed do. This generator builds you skill number one.