TAO DAFT Dataset Converter

The skill declares only Read and Bash and scopes inputs and outputs to local DAFT dataset paths; no malware, credential theft, or covert exfiltration is evident. Points are deducted because it installs an external Python package, executes conversion, and writes to an output directory without explicit user confirmation, overwrite protection, rollback, sensitive-data handling, or complete data-flow disclosure.

The workflow is internally coherent: it checks the version, discovers nested --help surfaces, and includes several common-error diagnostics. Points are deducted because no committed tests cover the conversion paths, the dependency is not version-pinned, and the claim of cross-version stability lacks static support; the static calibration therefore caps this at 10.

Trigger conditions, DAFT versus non-DAFT boundaries, input granularities, and target-format discovery are reasonably clear. Points are deducted for limited non-fit coverage, no explicit user-output contract or Chinese-language support, and no evidence that dependency installation is reachable from mainland-China networks; the core function is local CLI work and does not entirely depend on overseas services.

The material provides a name, purpose, prerequisites, version, Apache-2.0 license, limitations, troubleshooting, and publication-evaluation information, with generally readable structure. Points are deducted for missing Examples and FAQ sections, no changelog or explicit maintenance/update path, a nonconforming author format noted by the benchmark, duplicated quick-start material, and partially inconsistent license wording between the skill card and repository metadata.

The skill turns the conversion request into an actionable sequence of version checking, help discovery, source/target selection, execution, and post-conversion validation. Points are deducted because no committed real conversion examples or output artifacts demonstrate correctness, and comparative benefit or alternatives are only lightly documented; static calibration prevents a score above 7.

Versioned documentation, an evaluation report, one positive evaluation task, and CI/publication governance files provide auditable evidence for routing and documentation behavior. Points are deducted because the evaluation covers planning rather than actual conversion, with no conversion test suite, independent corroboration, or reproducible end-to-end evidence; static calibration prevents a score above 5.