publishing-agent-skills-safely
Use when preparing an agent skill for public distribution and a repeatable safety, packaging, and provenance check is needed.
SKILL.mdskills/publishing-agent-skills-safely/SKILL.md
GitHubPublishing Agent Skills Safely
Package a public agent skill only after rebuilding its examples with synthetic data and running the local release checks.
What this skill does
- Validates a candidate skill with a public-safe leakage scanner.
- Demonstrates an expected failure against an intentionally unsafe temporary fixture.
- Creates a deterministic ZIP package from a valid synthetic fixture.
- Produces a machine-readable publication report for review.
Required inputs
- A public skill directory containing
SKILL.md,skill-release.json, and.skillignore. - Optional external denylist file held outside the public repository.
Workflow
- Run
python scripts/check_prereqs.py. - Run
python scripts/run_demo.py --output-dir demo/outputto verify the synthetic reference fixture. - Scan the candidate skill and its ZIP with
python tools/leakscan.py <path>from the repository root. - Review every finding. Do not suppress a finding by masking a name; rebuild the sample with generic terminology and synthetic data.
- Package only after the source tree, ZIP, and generated report are clean.
Safety rules
- Use reserved domains such as
example.comand synthetic identifiers such asDEMO-100. - Keep real organization, customer, ticket, schema, URL, email, screenshot, generated artifact, and credential data out of public fixtures.
- Keep any organization-specific denylist outside version control and pass it with
--private-denylistduring release verification. - Do not place credentials in configuration, examples, shell history, or package artifacts.
Demo
The demo validates a clean synthetic fixture, creates a temporary unsafe fixture at runtime to prove detection, packages the clean fixture, and writes publication-report.json.
Validation
Run the skill tests from the repository root:
python -m unittest discover -s skills/publishing-agent-skills-safely/tests -t skills/publishing-agent-skills-safely -v