git:20260617.174aad8 to git:20260713.d210701

7 added, 22 removed. Audit A to A.

---

name: best-practices

- description: "Use when the user wants state-of-the-art knowledge, best practices, or evidence-based analysis of ANY topic — technology, health, fitness, nutrition, finance, design, psychology, architecture, productivity, learning, or any field where humanity has accumulated knowledge. Use when user says best practices, state of the art, most effective, optimal, evidence-based, how should I, what's the best way to, pros and cons, comparison, or /best-practices."

+ description: "Research current best practices and turn the evidence into a clear recommendation. Use for state-of-the-art, evidence-based, optimal-approach, comparison, or pros-and-cons questions in any field."

argument-hint: "<topic>"

user-invocable: true

- allowed-tools: WebSearch, WebFetch, Read, Glob, Grep, Agent, TaskCreate, TaskUpdate, TaskList

+ allowed-tools: WebSearch, WebFetch, Read, Glob, Grep, Skill

---

- # Best Practices — Universal State-of-the-Art Research

- You are an epistemic filter, not an encyclopaedia: surface accumulated domain wisdom by gathering the strongest sources, grading their evidence, and synthesizing an honest, actionable report.

- No `$ARGUMENTS` → ask the topic and STOP.

+ # Best Practices

- ## Process

+ Match the research effort to the decision: use a quick lookup for a narrow fact; use deep, multi-source research for costly, risky, disputed, or fast-changing choices.

- 1. **Domain profile.** Classify the topic's field, its evidence gold standard, its rate of change, and the variables the answer depends on (experience, goals, constraints, time horizon). If critical context is missing, ask or give conditional recommendations.

- 2. **Mandatory sources when the topic is Claude Code** (skills, hooks, subagents, CLAUDE.md, plugins, MCP, agentic coding) — fetch all in parallel:

- - `~/Documents/ai/superpowers` (Read local; fallback WebFetch GitHub)

- - `~/Documents/ai/claude-code-best-practice` (Read local; fallback WebFetch GitHub)

- - WebFetch `https://github.com/anthropics/claude-code/releases`

- - WebFetch `code.claude.com/docs/en/best-practices`, `/skills`, `platform.claude.com/.../agent-skills/best-practices`

- 3. **Gather** evidence with domain-adapted web searches (consensus, trade-offs, studies, failure modes/myths, domain-authority sites, implementation, frontier). Run in parallel via Agent subagents. For named domain-authority anchor sources, use `references/research-strategy.md`.

- 4. **Assess.** Tag each major claim `[strong]` / `[moderate]` / `[emerging]` / `[contested]` / `[insufficient]` / `[convention]`. Map where authorities agree, disagree, and where evidence is absent.

- 5. **Synthesize** for an intelligent non-specialist. Carry the evidence tags through, attach validity envelopes to each practice, give a decision framework where multiple approaches are valid, and flag where experts disagree. Match the user's language, cite inline, be opinionated where evidence is strong and honest where it is weak.

+ On Claude, use `/deep-research` when available. On Codex, perform the equivalent iterative web research. Otherwise use the platform's best search and browsing tools.

- ## Quality gates (drop silently on failure)

+ Prefer primary sources, cite material claims inline, and separate evidence from inference. Explain consensus, disagreement, uncertainty, and where each recommendation applies. Recommend for the user's goals and constraints, not for an imaginary average user.

- - Verifiable domain authority, not marketing/AI-summary/anecdote/overclaim.

- - Not older than 4 years unless foundational/seminal (mark as historical).

- - Tech repos: ≥1,000 stars AND active within 6 months (both required).

- - If gates leave a section empty, write "No sources met the quality bar" — never lower the bar.

+ If missing context changes the answer, ask for it or give clearly conditional recommendations. Do not pad the report with source quotas, evidence labels, or weak secondary material.