Skill Card Generator
Generates or refreshes governance cards for existing agent skills.
The skill declares read/write scope for the target directory, references/, scripts/, and /tmp/, forbids credential and unrelated-file access, and restricts shell execution to three scripts. Human review before submission reduces release risk. It lacks explicit user confirmation, rollback, dependency-integrity checks, and detailed data-flow disclosure, so 7 points are deducted.
The workflow order, script roles, prerequisites, failure causes, and troubleshooting guidance are documented, and the benchmark reports security, correctness, and discoverability results. Static review cannot reproduce key paths; jinja2 is unpinned and abnormal-input coverage is not demonstrated, so the score remains below the static ceiling.
Activation criteria, non-fit scenarios, arguments, and intended outputs are reasonably clear, and the core function is local file processing rather than dependence on an overseas online service. Chinese-language interaction, runtime differences across agents, and finer semantic trigger boundaries are not documented, so 3 points are deducted.
The documentation provides progressive structure, permissions, dependencies, script arguments, examples, limitations, troubleshooting, licensing, and an evaluation report. The skill lacks an explicit version, changelog, maintenance owner, and update path; the supplied license metadata is NOASSERTION despite license declarations in the files, leaving a governance-verification gap and costing 5 points.
The stated goal is to produce a renderable and validateable governance card, with defined JSON context, Markdown output, and a submission gate; the committed 11-task benchmark reports favorable correctness and effectiveness results. Independently verifiable representative artifacts and reproduction are absent, so 1 point is deducted under static calibration.
SKILL.md, scripts, the style guide, evaluation tasks, and benchmark results provide auditable material and distinguish human-review markers. There is no CI, raw execution log, committed output artifact, or independent corroboration, so evidence coverage is limited and the score is kept below the static ceiling.
- No explicit red-line risk was found, but per-operation user confirmation and rollback guidance are missing before writing to the target directory or /tmp/.
- The jinja2 dependency is unpinned, with no documented trusted source or integrity-verification method.
- BENCHMARK.md contains summary results rather than raw logs, CI configuration, or reproducible artifacts; it should not be treated as independent validation.
- License governance has an unresolved verification discrepancy between CC-BY-4.0 AND Apache-2.0 declarations, SPDX notices, and the supplied NOASSERTION metadata.
- Chinese localization and compatibility with mainland-China agent/network environments are not documented.
What does this skill do, and when should you use it?
Skill Card Generator creates draft NVIDIA governance cards for existing agent skill directories. It reads permitted skill files and references, extracts structured signals, builds a JSON context, and renders a Markdown card. A validation script checks that review markers have been removed before submission. The result remains subject to human legal, safety, and owner review.
Reads SKILL.md and permitted files from a target skill directory, runs discover_assets.py for structured signal extraction, builds a context JSON file, renders a Markdown skill card with render_card.py, and checks for remaining VERIFY or SELECT markers with validate_submission.py. File writes are limited to the target directory and /tmp/, and shell execution is limited to three listed scripts.
- A skill owner creating an NVIDIA governance card for an existing skill.
- A maintainer refreshing a card after the skill changes.
- A team preparing legal or safety review material for a skill.
- A reviewer checking that manual-review markers are resolved before submission.
What are this skill's strengths and limitations?
- Defines a concrete discovery, rendering, and pre-submission validation workflow.
- Restricts permitted file reads, writes, and executable scripts.
- Uses structured JSON context and deterministic Markdown rendering.
- Clearly requires human legal, safety, and owner review.
- Only targets existing agent skill directories, not models, datasets, containers, or full systems.
- Requires Python 3 and jinja2.
- Validates context shape but not the legal or safety correctness of field values.
- The supplied source does not document a test suite, supported-platform matrix, or failure-recovery mechanism.
How do you install this skill?
Install it from the NVIDIA skill collection with: npx skills add nvidia/skills --skill skill-card-generator --yes. The README does not document a separate manual-copy installation procedure for this skill.
How do you use this skill?
Provide an existing skill directory as the target, for example: “Generate or update a governance skill card for /path/to/target-skill.” The directory must contain SKILL.md or skill.md. The skill then reads permitted files, runs the listed Python scripts, and produces a card; the exact output path depends on the caller’s context and output arguments.