agent-skills
Original, platform-neutral agent skills — prompt assets that work across Claude Code, Codex, Cursor, OpenCode, Windsurf, Gemini, GitHub Copilot, CodeBuddy, and Hermes.
No runtime dependencies, no network calls, no build step. Each skill is a self-contained folder of instructions; where a skill supports multiple platforms, the platform files are thin adapters that route every agent to the same shared behavior.
Skills
| Skill | What it does |
|---|---|
| diagram-generator | Turns architecture briefs and codebase notes into standalone HTML/SVG technical maps: system, implementation, infrastructure, trust, data movement, deployment, and process views. Full package: generation rules, planning IR, HTML template, validation script, per-platform smoke tests. |
| clarify-before-acting | A behavior rule for non-trivial requests: do your own homework first, reuse answers already in memory, work a question queue one item at a time until every direction-changing unknown is resolved, and never guess where direction is at stake. |
| insight-challenge | Adversarial verification of code-review findings. Challenges each claim with counter-evidence anchored to actual source code and emits structured verdicts — overturned / weakened / confirmed — instead of trusting findings as-is. |
| workflow-skill-designer | Meta-skill that designs orchestrator + phases workflow skills: a thin SKILL.md coordinator plus progressively loaded phases/ files, with TodoWrite patterns, inter-phase data flow, and validation checklists. |
| defect-detective | The first stop before fixing a bug — read-only triage for a defect signal (bug report, stack trace, or log) that answers: is this a known issue, who owns it, and what changed. Normalizes a stable fingerprint, searches the tracker for duplicates cross-stack, correlates recent changes when the implicated layer is known, and infers the owning team with an auditable evidence chain. Upstream of fix-it skills (triage, not repair); strictly read-only, with a ready-to-paste ticket draft but never files or mutates anything. Smoke-tested against a real Jira tracker (dedup hit and no-match paths) plus git blame correlation. |
| session-retrospective | End-of-session retro that logs development-workflow issues (wrong answers, skipped gates, wrong routing, MCP errors, skill/doc gaps, multi-turn corrections) as enum-validated JSONL signals — a recurring issue is a rule gap and its priority escalates; when nothing went wrong, one success entry, never fabricated. One phrase enters Fix Mode: items ordered by recurrence and priority, per-item proposals, one confirmation per item with an append-only audit trail, running from the log alone across sessions. Ships a zero-dependency validation script (Node 18+) that machine-checks log health and recurrence groups. The signal-data layer: machine-aggregable output, complementary to report- and memory-style retro skills. Smoke-tested on a real session across both the main path and the corrupted-line exception. |
Install
Each skill is standalone — copy or symlink only what you need.
Claude Code
ln -s /absolute/path/to/agent-skills/skills/<skill-name> ~/.claude/skills/<skill-name>
Codex (skill mode)
ln -s /absolute/path/to/agent-skills/skills/<skill-name> ~/.codex/skills/<skill-name>
Other agents
- diagram-generator ships per-platform adapters:
AGENTS.md(Cursor / OpenCode / Windsurf),GEMINI.md,COPILOT-INSTRUCTIONS.md,CODEBUDDY.md,HERMES.md— see its README. - The other five skills are Claude-skill-format folders; on platforms without a skills mechanism, point the agent at the skill's
SKILL.mdas an instruction file.
SkillHub (Community Marketplace)
Install any skill directly from SkillHub:
# Install the CLI once
curl -fsSL https://skillhub.cn/install/install.sh | bash
# Install a skill (namespace: user_95900c57)
skillhub install yumingtao-diagram-generator --namespace user_95900c57
skillhub install yumingtao-clarify-before-acting --namespace user_95900c57
skillhub install yumingtao-insight-challenge --namespace user_95900c57
skillhub install yumingtao-workflow-skill-designer --namespace user_95900c57
skillhub install yumingtao-defect-detective --namespace user_95900c57
skillhub install yumingtao-session-retrospective --namespace user_95900c57
ClawHub (OpenClaw official registry)
Install any skill from the ClawHub global registry:
# Install the CLI once
npm i -g clawhub
# Install a skill (owner: yumingtao)
clawhub install yumingtao/diagram-generator
clawhub install yumingtao/clarify-before-acting
clawhub install yumingtao/insight-challenge
clawhub install yumingtao/workflow-skill-designer
clawhub install yumingtao/defect-detective
clawhub install yumingtao/session-retrospective
skills.sh (decentralized GitHub index)
No sign-up required — skills are indexed directly from this public repository. Install via the skills CLI:
# Install all six skills at once
npx skills add yumingtao/agent-skills
# Or install a single skill
npx skills add yumingtao/agent-skills --skill diagram-generator
npx skills add yumingtao/agent-skills --skill clarify-before-acting
npx skills add yumingtao/agent-skills --skill insight-challenge
npx skills add yumingtao/agent-skills --skill workflow-skill-designer
npx skills add yumingtao/agent-skills --skill defect-detective
npx skills add yumingtao/agent-skills --skill session-retrospective
DeepSeek Harness (DSH)
Each skill ships as its own npm bundle, so you can install only what you need. Pick one of the two methods below.
Method 1 — npm packages (recommended; per-skill install)
dsh plugin --profile web add yumingtao-diagram-generator
dsh plugin --profile web add yumingtao-clarify-before-acting
dsh plugin --profile web add yumingtao-insight-challenge
dsh plugin --profile web add yumingtao-workflow-skill-designer
dsh plugin --profile web add yumingtao-defect-detective
dsh plugin --profile web add yumingtao-session-retrospective
On first boot of the web profile each bundle copies its skill directory into
~/.agents/skills (the user-global skill root shared by DSH and Codex CLI).
Method 2 — market index (add all six at once)
Add the market index, then install any skill. Pick one index URL:
# Option A — jsDelivr CDN (recommended; may cache `@main` for minutes to hours)
dsh plugin index add https://cdn.jsdelivr.net/gh/yumingtao/agent-skills@main/docs/market/index.json
# Option B — GitHub raw (always fresh, no CDN cache)
dsh plugin index add https://raw.githubusercontent.com/yumingtao/agent-skills/main/docs/market/index.json
# Option C — local file (for offline / development)
dsh plugin index add file:///path/to/agent-skills/docs/market/index.json
Install:
dsh plugin install diagram-generator
dsh plugin install clarify-before-acting
dsh plugin install insight-challenge
dsh plugin install workflow-skill-designer
dsh plugin install defect-detective
dsh plugin install session-retrospective
With the market index, skill content is fetched from
github.com/yumingtao/agent-skillsat the pinned commit recorded in the index, regardless of which index URL you use. The CDN/raw choice only affects how the index file itself is delivered.
Sync
sync/ is a generalized "one source → many platforms" distributor for the whole collection. A platform target receives the canonical skill folder as a symlink (default) or as a materialized copy.
sync/apply.sh --dry-run # preview
sync/apply.sh # link into every detected platform
sync/apply.sh --platform claude --skill insight-challenge # narrow scope
sync/apply.sh --to ./ --platform cursor # project-local install
sync/check.sh # verify nothing drifted
The distribution map lives in sync/manifest.sh — see sync/README.md.
Design principles
- One source of truth per skill. Platform files route; they never restate behavior.
- Markdown-first. A skill is instructions plus optional specs/resources — nothing to install or run.
- Progressive loading. Coordinators stay thin; execution detail lives in
phases//specs//references/loaded on demand.
Roadmap
-
drawio-firstoutput mode fordiagram-generator(needs dedicated mxGraphModel rules, a template, and import validation)
License
MIT. Skills in this collection are released under the same MIT terms.
No comments yet. Be the first to write one.