Instruction changes need behavioral regression
Google treats a skill as an artifact that changes tool choice and execution, not simply documentation. A phrase can induce an invalid flag or a loop, so valid Markdown is insufficient. Behavioral changes require version bumps; existing skills are evaluated against their previous version and new skills against an agent without the skill.
Fast structural, frontmatter, naming and version checks precede expensive evaluations and AI review. Reports must match the latest commit. Version strings and proper major/minor resets support installation, pinning and comparison; they do not imply a network update on every invocation.
Atomic criteria and executable steps
The article illustrates a storage-bucket expectation combining region, access control and lifecycle rules. Missing one part produced inconsistent judging. Three independent criteria make failures attributable. Validation also checks nonempty prompt files, duplicate prompts, conflicting expectations and stale reports.
Deterministic command sequences belong in scripts while judgment stays in instructions. Reconstructing rigid command chains, JSON parsing or deployment steps probabilistically creates avoidable variation. Scripts use noninteractive inputs, structured output and actionable errors. Compact core instructions link directly to detailed references, avoiding chains that agents may stop following halfway through.
Transferable process lessons
Overbroad path rules once blocked typo fixes while missing newly added skills in deeper directories. Gates therefore need operation-aware path coverage. Generated indexes prevent catalog drift; AI review checks design problems before subject-matter approval.
My interpretation is to separate valid files, current evaluation evidence and worthwhile behavior. A version bump records a change but does not prove its quality. The article does not establish that every skill improves performance, and repository popularity is not evaluation evidence.
Source: Original engineering article