Skip to content
Nash Software Services
NASHSOFTWARE SERVICES
← Writing

Building a Capture Learnings Skill

How I built a skill that captures and preserves learnings from each development session.

AIClaude CodeWorkflow

I added one command to the end of every Claude Code session. Now it actually remembers what we learned.

One of the things that bothered me early on was this: you spend an hour in a debugging session, you work out exactly why something was going wrong, you find some niggly Windows-specific path issue or some TypeScript pattern that keeps biting you, and then the session ends and Claude just... forgets. Next session, you're starting from zero. You might even go around in circles with the same problem two weeks later because nothing got captured.

So I built a skill for it. I call it /capture-learnings, and I run it at the end of any session where things got complicated - debugging sessions, sessions where a lot of new context came in, sessions where something went wrong and we had to work out why.

What it does is extract everything useful from the session and persist it to the right place. Two buckets: global tech knowledge, which goes into files like ~/.claude/memory/tech/node-windows.md or typescript.md depending on what the session was about, and project-specific stuff, which goes into the local project memory directory. The global stuff is patterns that apply across any project - "this is how Windows handles file paths differently", "this is why that TypeScript config option matters" - and the local stuff is rules specific to that codebase.

It also identifies skill gaps - recurring manual workflows from the session that might benefit from becoming a skill - and outputs ready-to-paste /skill-creator prompts for any it finds. So you're not just capturing knowledge, you're building the automation on top of it too.

Runs on Haiku because it's memory consolidation, not reasoning - fast and cheap. And it launches a background agent to do the actual work, otherwise on a long session it was just eating context.

Honestly it's one of those things that's a bit boring to set up, but once it's in the workflow you stop noticing it and you just notice that Claude is getting better over time - actually keeping up with the project instead of starting fresh every session.

Drop a comment or follow if you want the SKILL.md file.

(Genuinely asking - is it more useful as a LinkedIn message, or should I put it in a repo? Not sure which is better here.)

Share

Interested in bringing this kind of AI-assisted delivery to your team?

See AI & Automation

Related posts