Found something good?

Self-improving Claude Code skills are f*cking ridiculous 🤯

One loop → 10 test runs, scored against an eval, prompt rewritten, retested, winner kept.

A hook writer skill went from 32/50 to 47/50 overnight.

All inside Claude Code.

Perfect for DTC brands and agencies who have built Claude Code skills but the output is still inconsistent — great 70% of the time, unusable the other 30%.

If you've been manually tweaking your skill prompts one run at a time, re-reading outputs, adjusting instructions based on vibes, and never quite getting the consistency you need...

This method eliminates the entire loop:

→ You define 3-5 binary eval criteria for your skill
→ Claude runs the skill 10 times with varied inputs
→ A separate evaluator scores every output against your criteria
→ It identifies the most common failure patterns
→ Rewrites the skill prompt to fix what's failing
→ Retests and keeps the winner
→ Repeats until the score plateaus

No manual prompt tweaking.
No reviewing every output by hand.
No "it worked that one time but I can't reproduce it."

What you get:

→ A skill prompt that's been through 50+ automated test runs
→ A scored improvement log showing exactly what changed and why
→ Eval criteria you can reuse every time you update the skill
→ A method that works on any skill: hooks, briefs, ad copy, scripts, reports

Inspired by @karpathy's auto research repo, the same loop AI labs use to improve their own models, applied to your creative workflow.

I put together a full playbook showing how to set up the eval, the exact Claude Code prompt for the improvement loop, and starter eval criteria for the 5 most common DTC creative skills.

Want the playbook for free?

> Like this post
> Comment "IMPROVE"

And I'll send it over (must be following so I can DM)