What’s up!

Last week, I found the optimize_anything paper and shared it with Sugi.

Sugi got inspired (read: nerd sniped) and wrote an amazing skill that lets Claude Code rip through this type of optimization.

You give it any document — a prompt, a spec, a pitch email — and it runs iterative feedback loops until it's actually good.

Works like RL-style refinement, but the judge can evaluate almost anything and give specific feedback (not just a score). So it converges in 3-5 rounds instead of thousands.

Read and test it out here, Simmer: A Self Honing Skill.

Hit us up with any questions,
Harper and the 2389 team :)