Standing instructions (CLAUDE.md template)

A generic, reusable CLAUDE.md skeleton for Claude Code. Fill the About / projects / infra blanks with your own context, delete what you don't need, and keep it short — a bloated CLAUDE.md gets ignored (Anthropic's own guidance). The real power lives in the skills/ (loaded on demand), so this file stays lean and points at them.

About me / my projects

• <Who you are, your role, the stack you work in.>
• <Your main repos + where they live + how they deploy.>
• <Any persistent context an assistant should read at session start.>

At the start of every non-trivial prompt

1. Interview me first — get a full download before building. Use AskUserQuestion (2–4 options, multiSelect when relevant) before assuming requirements. Surface the ambiguities I might not have spelled out across: scope (in/out), target environment, output shape, tradeoffs (speed vs thoroughness / cost), done-criteria. Recommend a default per question so it's a confirmation, not a blank. (Anthropic's own "interview-first spec" recommendation — see the working-with-claude skill.)
   • Skip only for one-line lookups, single trivial commands, or when I answered the same question this turn.
2. Plan with a task list for any task with 3+ distinct steps; execute step-by-step, mark each done inline.
3. Sync first before editing a repo: git status && git log --oneline -5 && git log --oneline -5 origin/main; rebase if behind. (Repos with auto-writers drift constantly.)

Working with Claude (operating principles)

Load the working-with-claude skill for the full playbook; the always-on rules:

• Explore → Plan → Code → Commit. Read the relevant code first; for 3+ step work, plan before coding. Always hand work a verification check (test/typecheck/build/live URL) so it self-checks.
• Context is the bottleneck. Use subagents for investigation (they return tight summaries); /clear between unrelated tasks; keep this file lean.
• Match the machinery to the task. An agent ≈ 4× a chat turn's tokens; multi-agent ≈ 15×. Fan out / run a workflow only when the task value justifies it. Don't multi-agent a trivial task.
• Research locally first: inspect local code/logs/config before the web; browse only for external facts.
• Cost/speed defaults: stable prompt prefix first (cache reads ≈ 0.1× base); cheap model for noisy/bulk, mid for default, top tier for reasoning/agentic; cap output, ask concise; parallel tool calls in one message.

How to communicate with me

• Be terse; no trailing "summary of what I did". Preserve raw tool output when it matters. Match response length to task complexity. No emojis unless asked.

How to delegate to subagents

• Brief them like a colleague who just walked in (no session context). Structure prompts with XML tags (<task>, <context>, <constraints>, <output_format>). Cap response length. Pick model by task (top tier for synthesis/review, mid for implementation, cheap for noisy work). Run independent subagents in parallel (multiple tool calls in one message).

How to ship code

• Pre-commit gate: run lint + types + tests BEFORE every commit (the block-no-verify hook enforces it).
• Commit + push promptly; push-then-tag, never tag-then-push. For UI/web changes, verify the live deploy.
• Before marking a task complete: confirm tests pass AND the live surface reflects the change.

Design (load before any web/UI work)

• Before building/restyling any web UI, load the design-system skill first — it routes to the matching per-type design skill and supplies tokens, fonts, a11y, components, motion, and a premium-feel checklist.

Safety

• Never paste secret values in chat (the secret-scan hook warns if you do). Use env files.
• Confirm before destructive actions (rm -rf, force-push, hard reset, dropping data, killing prod).
• Don't weaken lint/type/CI configs silently (the config-protection hook asks first).