An Agent Skill that holds an AI coding agent's output to the standard of current best practice — the recommended, idiomatic way now, not merely an answer that runs — by teaching it when to distrust its own (stale, version-fuzzy) memory and verify against authoritative sources.
中文说明请见：README.zh-CN.md
Training knowledge has a cutoff and is fuzzy about versions. In fast-moving areas — frameworks, SDKs, toolchains, "best practices" — an agent will be confidently wrong and usually won't notice. But "search every time" is an anti-pattern too: it pollutes context, slows everything down, and pulls in low-quality content.
This skill is pure judgment, with zero new infrastructure. It teaches two things models get wrong:
- Predicting your own blind spots — version/time-sensitive facts, "idiomatic way" questions, code against external APIs.
- Recovering after being corrected — the moment the user corrects you, doubts you, or repeats a request is ground truth that your memory was wrong. Stop re-guessing from the same memory. Verify first.
It deliberately does not build a knowledge base or cache results — cached version-sensitive facts go stale and then get re-trusted with false authority. Verify when it matters, then move on.
Install as an Agent Skill — tell any AI coding tool that supports agent skills (Claude Code, Codex, OpenCode, Cursor, Antigravity, etc.):
Install the skill from liustack/best-practices
Or use the skills CLI directly:
npx skills add liustack/best-practices --skill best-practicesOr install manually by copying the skill into your agent's skills directory:
# Claude Code (personal, global)
cp -r skills/best-practices ~/.claude/skills/best-practices- When to distrust yourself — internal signals (version-sensitive facts, fast-moving domains, "best practice" questions, code against external APIs) and external signals (user corrects/doubts/repeats — highest weight).
- The doubling-down rule — once corrected, stop answering from memory. Verify, then speak.
- Triage before reacting — being corrected tells you something is wrong, not what. Re-read context (cheap) before reaching for the web.
- Guardrails — don't over-verify stable fundamentals, locally confirmable facts, or things the user already specified.
- Source hierarchy — installed source / lockfile > official docs (MCP /
llms.txt) > web search, always version-aligned.
MIT