Improve a skill until it clears a score: tessl review fix --threshold 90 --max-iterations 5. A find-fix-re-verify loop (pattern-autonomous-loop) whose target is a capability rather than a product — the agent rewrites the instructions, re-scores them, and repeats until the rubric is satisfied or the iteration budget runs out.
Small, and worth a page for what it implies. Once a capability has a machine-readable quality score, improving it becomes an optimization problem an agent can run unattended — the same shape as gstack-qa’s self-fixing QA loop or ce-dogfood, pointed at the skill corpus. The bounded iteration count and the required confirmation prompt (waived with a flag) are the guardrails on a loop that is, after all, rewriting the instructions every future agent will follow.
The obvious caution is that it optimizes for the rubric, not for effect. tessl’s own framing keeps the two apart: a review checks form, an eval checks whether behavior changed. A skill hill-climbed to 95 on structure that tessl-eval shows changes nothing is a real outcome, and the framework-side instinct against exactly this failure is pstack-principle-test-behavior-not-implementation.
See Also
- tessl-review — supplies the score this climbs.
- tessl-eval — the check that the climb was worth anything.
- pattern-autonomous-loop — the technique, aimed at a capability.