A Smaller Skill Still Has to Preserve the Contract

Agent skills tend to grow by addition. A workaround becomes a paragraph, a paragraph gets copied into several branches, and the artifact becomes harder to inspect. Shortening it can help, but only if the shorter form preserves the governed behavior.

WTK has already asked whether a skill graph beats direct retrieval and whether a whole-task skill transfers to later work. This note is not about selecting or transferring a skill. It starts after one skill has been chosen and asks whether its representation can shrink without erasing a rare contract obligation.

The signal

SkillZip represents a skill as typed contract units plus residual content. Its optimizer looks for reusable structure while requiring coverage of extracted triggers, workflow edges, tool requirements, obligations, and output fields. Its continual mode applies the same idea when a new patch is accepted.

For WTK, the proposed mechanism is not automatic compression. It is a coverage-preserving candidate representation that still has to beat the current skill on held-out behavior.

The evidence

The paper evaluates one-shot compression on BFCL-v4 Web Search, BrowseComp, and AIME 2025, then studies continual compression over self-evolution rounds. It reports that structural compression reduces skill length while retaining benchmark performance better than textual reduction baselines. The online design uses a write-ahead log, validates schema and extracted-unit coverage, and replaces persistent state only after rendering succeeds.

That design contributes a useful engineering pattern: explain a repeated rule once, reference it from its scoped uses, and keep unique or uncertain material as residual content rather than forcing it into an abstraction.

The boundary

Coverage is defined over units extracted by the method. It cannot prove that the extractor recognized every semantic obligation. The benchmarks do not exercise WTK authority boundaries, negative paths, target projections, or publication and deployment controls. “Evaluation-free” describes the compression procedure, not evidence that downstream qualification can be skipped.

The builder impact

A WTK skill consolidation should preserve source identity, every declared trigger, permission boundary, refusal path, output obligation, and provenance link. The admission question is semantic preservation under compression, not retrieval quality or cross-task utility. A smaller rendered skill is a candidate artifact, never replacement authority by itself. Rare safety clauses deserve explicit fixtures because their low frequency makes them easy to compress away unnoticed.

The WTK test

Compare one current WTK skill with a structurally consolidated candidate. Freeze the source digest, target harnesses, model bindings, budgets, positive and negative fixtures, and evaluator. Add mutation fixtures that remove a rare denial, mis-scope a trigger, merge incompatible tools, or lose an output field.

Measure artifact size, declared-unit coverage, held-out task completion, negative-path preservation, authority violations, target parity, maintenance cost, and catalog outcome. The candidate succeeds only if it is materially smaller while all contract, denial, and qualification checks remain at least as strong. Any lost obligation or false authorization is a failure. No useful size reduction with preserved behavior is inconclusive. This is a proposed WTK experiment, not authorization to rewrite a skill.

Still unknown

We do not know whether WTK skills contain enough repeated structure for this approach to outperform careful manual consolidation.