Last updated: August 17, 2026
Maestro: Ablate is a general-purpose audit and measurement tool for Agent Skills. It ranks skills by context cost, audits a SKILL.md against a written rubric, rebuilds the file, and runs a measured ablation loop against a model the operator selects.
Two limits are inherent to the method and are stated in the skill's own documentation. Absence of harm is not presence of benefit: a passing ablation shows that a cut did not break the tasks that were run, not that the cut improved anything. Results are specific to the model and CLI version measured against: a skill trimmed against one model may under-serve the next. Reports record both for that reason, and version control is what allows a skill to be re-expanded.
Ablate output is not proof that a rebuilt skill is correct, complete, safe, compliant, or better than the original. A passing gate, a score, a diff, or a grader verdict does not establish production readiness.
Users are responsible for version-controlling skills before rewriting them, reviewing every proposed cut, budgeting the model spend a Phase B loop incurs on their own account, deleting labs that hold copied credentials, and testing rebuilt skills in real work. Independent review may be required where a skill governs legal, regulatory, medical, financial, or safety-critical work.
The MIT License governs permission to use, copy, modify, and distribute this software. The software is provided "as is" without warranty. The warranty and liability terms in the MIT License apply.