Professors at the University of Illinois just dropped a 22-page PDF on agents that write their own Claude Code skills.

generate a skill → a separate verifier writes its own tests → fix what fails → repeat

stop hand-writing skills.

the agent writes them and checks itself.

one agent builds the full multi-file skill, not a single function.

a second agent, blind to the code and the real tests, invents its own checks.

the verifier writes harder tests instead of leaking the answer.

it rewrites itself until it holds.

read the paper first, then the article below.