g0ld2k Skills
Focused personal skills for PR authoring, review feedback, and release notes, with small compatibility and on-demand tools. The harness owns general planning, delegation and branch orchestration; repository guidance owns project constraints and verification commands.
Skills
| Skill | Use |
|---|---|
adversarial-review | Challenge unnecessary complexity using simplicity and correctness perspectives. |
test-adequacy-review | Find low-quality tests worth strengthening, replacing, or removing, and missing behavioral coverage. |
pr-generator | Draft or publish the requested PR title/body from branch evidence. |
pr-comment-review | Triage and address existing PR feedback and scoped replies. |
testflight-notes | Draft tester-facing notes from a verified history range. |
commit-message | Draft copyable Conventional Commit messages from staged changes or an explicitly selected source. |
simplify | Optional focused review or requested cleanup. |
prune-comments | Remove low-value comments while preserving useful maintenance knowledge. |
pr-closeout-loop | Explicitly selected readiness or bounded closeout; monitoring and merging require their own scope. |
The three everyday skills do not automatically call every other skill. User intent and existing authorization take precedence over workflow preferences. See the 2026-10-04 simplification and PR disposition for retired entrypoints and preserved historical proposals.
Packaging
This repository is one self-contained Agent Plugins v1
plugin. plugin.json and skills/<name>/SKILL.md are the distributable source;
there is no generated skill tree. The Codex .agents/plugins/marketplace.json
and Copilot .github/plugin/marketplace.json adapters point at the root. The
manifest is validated against the pinned schema in schemas/agent-plugins/.
An explicit-only skill carries both Claude's disable-model-invocation: true
frontmatter and Codex's policy.allow_implicit_invocation: false in
agents/openai.yaml; neither client reads the other's guard. The validator
checks both for the explicitly selected closeout workflow.
_shared/conventions.md is vendored into its consuming skill directories by
scripts/sync-shared-conventions.py and drift-checked. This deliberate duplication
keeps a single-skill installation self-contained. Consumers are discovered from
entrypoint links to references/conventions.md.
Install
Direct single-skill example:
gh skill install g0ld2k/Skills skills/pr-generator/SKILL.md --agent codex --scope user
Codex marketplace:
codex plugin marketplace add g0ld2k/Skills --ref main
codex plugin add g0ld2k-skills@g0ld2k-skills
Copilot marketplace:
copilot plugin marketplace add g0ld2k/Skills
copilot plugin install g0ld2k-skills@g0ld2k-skills
Inspect skills before installation. Source changes do not themselves update an installed cache; record the installed source revision when diagnosing drift.
Contribute and validate
Use authoring guidance and focused evaluations. Add only resources that improve the actual workflow. Preserve authorization, target identity, unrelated work and truthful evidence.
Before reporting a skill addition or change ready, run the applicable existing
regression and packaging checks. The
validation workflow is the canonical
full command set. When adding, renaming, or removing a skill, update the expected
skill set in test_skills_are_discoverable_at_the_plugin_root in
tests/test_plugin_packaging.py.
Common local commands, run from the repository root:
python3 scripts/sync-shared-conventions.py
python3 scripts/validate-skills-repo.py
python3 -m unittest discover -s tests
gh skill publish --dry-run
CI also verifies shared-file drift, JSON/shell syntax and ShellCheck. The publisher command above validates only; it does not publish a release. Structural checks do not substitute for behavioral evidence or client/runtime testing.
Reuse passing results while the relevant files and environment remain unchanged; rerun affected checks after changes. Report what actually ran and any skipped or blocked checks. Label written or statically reviewed validation scenarios as such, rather than as executed tests.