“When you’re happy with the results, say: great, now review everything we’ve done and turn that into a skill. And then every time I use that skill, I’d say: learn from everything we did and update the skill. So it gets smarter each time you use it.” — Lou
Session context: 2026-05-28_Mastermind — Dirk faced a board-letter task he’ll repeat quarterly and asked how to make it repeatable without re-learning the whole process each time. Lou described a loop that turns a one-off conversation into a tool that improves itself.
Core Idea
A skill is usually treated as a fixed artifact: you build it once, then invoke it unchanged forever. Lou’s pattern treats the skill as a living thing that gets better every time it’s used. The loop has three moves. First, interview to build: have Claude interview you about the task (“ask me the questions I need to answer to write this”), go back and forth until the output is right, then say “review everything we’ve done and turn that into a skill.” Second, use it: run the skill on the next real instance of the task. Third — and this is the move most people miss — fold the corrections back in: as you fix and tweak the skill’s output, end the session with “learn from everything we did and update the skill.”
The result is a skill that compounds. Each use surfaces an edge case, a preference, a correction; each “learn and update” bakes that into the skill so the next run starts smarter. Over a few quarters, Dirk’s board-letter skill won’t just produce a generic business letter — it’ll have absorbed his audience’s preferences, the legal sensitivities, the structure his board actually reads. The first run is the worst the skill will ever be.
This works because the skill lives on disk in Claude Code: the skill file is editable in place, so “update the skill” is a real file write, not a suggestion that evaporates when the chat ends. It’s the same closing-the-loop discipline as the conversation audit, but pointed at a reusable tool rather than a single session’s decisions — every correction becomes permanent capability rather than a fix you’ll have to make again next time.
Practical Application
For any task you’ll repeat (quarterly reports, client follow-ups, proposals):
- Build by interview. Tell Claude what you’re trying to produce and say “interview me — ask the questions I need to answer.” Iterate until the output is right.
- Codify: “Review everything we did and turn it into a skill.” Include any example documents or a template so it has a model to follow.
- Use it on the next real instance.
- Close the loop: after you’ve corrected the output, say “Learn from everything we did this session and update the skill.” Watch the next run start closer to done.
The discipline is step 4. Skipping it means re-teaching the same lessons forever.
Related Insights
- Insight - The Conversation Audit Technique — Never Let a Session’s Fixes Evaporate — the same close-the-loop move applied to a session; this points it at a reusable skill instead.
- Insight - Skills Encode Judgment Into Persistent, Composable Intelligence — what the skill is accumulating each cycle is your judgment.
- Insight - Extend Claude With Skills to Build Your Personal AI Ecosystem — self-improving skills are the compounding unit of a personal AI ecosystem.
Evolution Across Sessions
Builds on Insight - The Conversation Audit Technique — Never Let a Session’s Fixes Evaporate (2026-04-09), which established capturing fixes at session end so they don’t evaporate. New development: this session extends that loop into a recursive self-improvement cycle on a reusable skill — “learn from what we did and update the skill” run after every use — so the tool compounds rather than just preserving a single session’s decisions.