Skill Creator adds evals, benchmarks, refinement

This is a meaningful step up from prompt tinkering to software-style skill development: Anthropic is making agent behavior something you can test, compare, and iteratively improve.

• –The new workflow adds four practical modes around skills: create, eval, improve, and benchmark
• –Evals use realistic prompts plus explicit assertions, which is much closer to regression testing than ad hoc prompting
• –Benchmarking against “with skill” vs “without skill” helps reveal when a skill adds real value versus when the base model already handles the task
• –Refinement matters because skills can decay as models change, so built-in measurement makes long-term maintenance much more realistic
• –For Claude Code users, this lowers the barrier to serious skill engineering by bundling the harness instead of forcing teams to build their own