agent-plugins

Cross-agent skills by KuaaMU. Plain SKILL.md directories — works in Claude Code, Codex, opencode, Cursor, and anything else that reads SKILL.md.

中文版

Install

# everything
npx skills add KuaaMU/agent-plugins

# just one
npx skills add KuaaMU/agent-plugins --skill mentor

The CLI writes to ~/.agents/skills/ and links or copies into whichever agent directories it finds. --agent codex narrows it, --global skips the project scope, -y skips the prompt. Run npx skills add --help for the rest.

Which skill do I need?

If you are…	Take
Not sure which skill fits — let the router decide	router (user-invoked)
About to claim "done / fixed / faster" — verify first	verify-first (model-invoked)
Want to actually understand the project while shipping, let AI mentor you	mentor (user-invoked)
Tackling a hard numeric target over days/weeks (pass rate, latency) and keep "making progress" without converging	long-horizon-skills
Running a long engineering or research mission and need drift-proof checkpoints	adaptive-mission
Turning finished work into lessons, briefs, or publishable artifacts	work-output
Building a private, cross-AI personal knowledge vault	personal-os

Skills

Skill	What it does
router	User-invoked skill router: when you don't know which skill fits, invoke it and it routes you (or tells you none fits). Also documents the repo-wide trigger policy.
verify-first	Model-invoked verification discipline micro-skill: 5-step loop (define evidence → real environment → raw output → discriminability check → seek counterexamples) before claiming anything works, plus a rationalization pre-buttal table.
mentor	User-invoked cognitive-alignment skill: a study companion that cross-session reads the main agent's work and mentors you — diagnosis-first, progressive guidance, 4 questions, 3 modes, Feynman checks. Never interferes with the main line. Core ideas: cognitive alignment + letting AI grow you.
adaptive-mission	Minimal-plan, drift-tolerant workflow for long engineering and research missions: 3–5 checkpoints, STATE/TRUTH/PLAN/REVIEW records, optional subagents, real-environment acceptance.
long-horizon-skills	Anti-failure + execution system for hard numeric targets over weeks: six failure modes, trigger-style rules, independent auditor, JSON state machine, racing discipline, file hygiene, hard-set, and domain profiles.
work-output	Distill any task into four-track outputs: deliverables, reproducible process traces, reusable lessons, and publishable artifacts.
personal-os	Cross-AI personal data vault methodology: private GitHub repo + AGENTS.md entry point + markdown memory any AI can onboard to.

Each is self-contained; install only the ones you want.

Adding a skill

No config needed: drop a directory with a SKILL.md under skills/ and add a row to the tables above. Keep name in the frontmatter identical to the directory name, or installers will skip it. Write the SKILL.md in English; declare user-invoked vs model-invoked in the description.

Quality gate (eval & pruning)

• Admission: a new skill must have name == directory name, a description stating its trigger + invocation mode, and a body ≤100 lines or an index structure (heavy content goes to references/).
• Eval: before/after a revision, blind-run the same real task and check whether agent behavior actually improved; text-only changes with no behavior change don't merge.
• Pruning: quarterly review — skills with no real usage or overlapping others get retired or merged; retirements are recorded in ARCHIVED.md (created on first use).

Notes

• long-horizon-skills ships seven runnable scripts (stdlib-only Python + one bash). Every script self-tests (--self-test); run them before merge.
• long-horizon-skills v2 (2026-10-10) merged the former standalone ascend-operator-tackling repo (now archived). Ascend/CANN specifics live on as references/profiles/ascend-cann.md.