Agent skills

English | 简体中文

Reusable Agent Skills for focused workflows. Each top-level directory is independently installable.

Choose the smallest set that covers the task. Project-specific skills take precedence over generic ones when both match.

Skill index

Skill	Purpose
grill-me	Stress-test a plan one decision at a time.
idea-capture	Save and maintain actionable ideas.
idea-next	Pick one evidence-based next action.
material	Capture verified fragments worth sharing.
draft	Turn fragments or raw input into publishable copy.
rewrite	Polish Chinese technical writing while preserving facts and voice.
repo-prose	Keep repository prose current, concrete, and free of authoring residue.
autoresearch	Prepare one measurable improvement experiment.
abtest	Prepare isolated control/treatment experiments.
simplify	Apply behavior-preserving cleanup.
verify	Verify public behavior with black-box evidence.
nanobot-webui-verify	Verify nanobot WebUI through the real gateway and browser.
nanobot-gate	Coordinate nanobot simplify, verification, review, and CI gates.
triage	Produce a short decision brief for complex artifacts.
herdr-scode-delegate	Delegate a background task to another scode session in Herdr.
pr-worktree	Maintain one attached worktree per branch.
pr-review	Review PR correctness, reachability, and merge value.
pr-fix	Apply an authorized focused fix to a PR branch.
pr-rebase	Rebase and verify a PR branch safely.
pr-label	Classify and update PR labels from repository evidence.
pr-metadata	Keep a PR title and description aligned with its pushed scope.

Routing

• Content: material -> draft -> rewrite.
• Repository prose: use repo-prose for documentation, comments, and contributor instructions; keep authorial publishing in rewrite.
• Agent delegation: use herdr-scode-delegate to assign a bounded background task to another scode session without supervising its pane.
• PR code work: use pr-worktree for local review, testing, fixes, and rebases. Use triage, pr-label, and pr-metadata directly against the remote PR; pr-fix composes pr-metadata after an authorized push.
• Verification: use generic verify across projects, nanobot-webui-verify for nanobot browser/gateway surfaces, and nanobot-gate for full nanobot readiness.
• Experiments: choose autoresearch for iterative improvement or abtest for control/treatment comparison.
• Human-readable plans and results stay in the conversation; skills persist deterministic state and raw evidence when useful.

External mutations such as pushes, PR comments, labels, merges, and remote writes begin from explicit authorization for that operation.

Installation

Clone the repository, then copy or link selected skill directories into the discovery path used by your agent:

git clone https://github.com/chengyongru/skills.git ~/src/agent-skills
mkdir -p ~/.agents/skills
ln -s ~/src/agent-skills/pr-worktree ~/.agents/skills/pr-worktree
ln -s ~/src/agent-skills/pr-review ~/.agents/skills/pr-review

Agent runtimes may use different discovery paths. Codex users can refer to Build skills.

Invocation

Explicit mentions are the clearest option when installed skills have adjacent scopes:

Use $idea-capture to save this idea.
Use $nanobot-gate to check this change before PR creation.
Use $pr-review to perform a read-only review of PR #123.

Contributing

• Give each skill one responsibility and precise trigger boundaries.
• Keep SKILL.md limited to non-obvious decisions, contracts, and helper usage.
• Put deterministic mechanics in scripts/ and genuinely optional domain knowledge in references/.
• Follow AGENTS.md for action-oriented language and direct verification output.
• Update both README files when skills or routing change.