Give it an article, talk, recent session, or rough idea. It pulls out the decisions and gotchas, then packages them as a skill your agents can run.
$npx skills add kylezantos/skill-distilleryTurn your agents into Neo.
I’m sure most of us remember that iconic moment when Keanu Reeves looks to his left at Laurence Fishburne and says “I know Kung Fu” (with an implied ‘dude’ in his vocal affect). He sits in a chair, gets a Smithsonian Library level of expertise loaded in, and walks out fluent. A SKILL.md file works like that for an agent. When constructed thoughtfully with reference files and structuring, it’s a little mini-app that loads a way of thinking, a process, or a framework into the model so it can execute like an expert.
So I built a skill that helps you make those little mini-apps in minutes. Paste in a URL of an article that describes a framework. Or a transcript of a podcast discussing a unique product approach. Maybe an idea for a system you thought of on the drive home, like one for managing your kid’s calendar. The skill pulls out the principles, the mental model, and the decision points, then packages them into something any AI agent can run.
It (largely) nails the first version because it stays interactive, clarifying intent with you while you build instead of running through rules and leaving you to fine-tune a misfire after. It even does another self-check review after it’s all built to make sure that the previous agents adhered to the proper taxonomy and construction.
I still surprise myself at times when I have an idea for a new type of one and it just... works. Like intent-engineering. A few of the other SKILL.md files I’ve made with it honestly feel like products masquerading as “little SKILL.md files.”
Different inputs, same bottom-up build.
The skill takes raw material wherever you have it and routes to the workflow that fits. An article and a half-formed idea route differently. Five entry points, same bottom-up synthesis underneath.
Hand it a URL, file, or pasted text. It reads the source, extracts the principles and mental models bottom-up, then designs the skill architecture with you. Articles, talks, podcasts all become skills you can interrogate.
/skill-distillery https://example.com/great-article-about-design-systems
Describe the skill you want. If the description is sharp, it proposes an approach. If it's vague, it brainstorms with you until the idea has edges. The blank-slate route, for when you know what you want but not what to feed it.
/skill-distillery
You just figured out a process with an AI and want to keep it. This mode pulls the repeatable workflow, the decision points, and the gotchas out of what you did, then packages them as a reusable skill. Process capture, not transcript dumping.
/skill-distillery capture
Point it at a skill you have or one you found. It runs the skill through a multi-lens evaluation (synthesis quality, architecture, interactive design, spec compliance, cross-agent compatibility) and gives you a rated assessment with prioritized improvements.
/skill-distillery audit ~/.claude/skills/my-skill
Extend a skill you've already built. Add a reference file, a new workflow, a script, or a template without rebuilding the whole thing. The skill knows what fits and what doesn't, so the additions don't drift from the original architecture.
/skill-distillery add reference
Works in any AI coding agent
One install command, any agent. The skill rides on the open skills protocol, so it works the same in Claude Code, Codex, Cursor, OpenCode, and 35+ others. The CLI auto-detects what you have installed and wires it up.
$npx skills add kylezantos/skill-distillery#Auto-detects your agent (Claude Code, Codex, Cursor, OpenCode, 35+ more)#Then run: /skill-distillery [URL | file | "description" | capture | audit <path> | add <component>]
Built from skills I kept rebuilding.
Skill Distillery is original work, but it didn’t come from nowhere. I leaned on Anthropic’s official skill specification for the format, and on the Compound Engineering plugin for the shape of a skill that builds skills. The interactive patterns (discovery phases, mode detection, wait gates) came out of skills I’d built and kept rewriting until they felt right.
The credit for the underlying ideas belongs to the writers and talks I keep going back to. This is the workbench I built to stop losing them.