rules-for-claude

rules-for-claude

My global Claude Code setup, mirrored from ~/.claude: the rules, the hook scripts that enforce them, and the global CLAUDE.md.

CLAUDE.md

The global instruction file, loaded at the start of every session before any project context. Three parts: hard coding standards (no type casts, no any, no unknown, no "pre-existing issue" excuses), a prose style guide that bans recognizable LLM writing patterns (performative contrasts, dramatic fragments, empty intensifiers, filler vocabulary, dashes as clause separators), and context-efficiency rules for file reading, subagent delegation, and response length.

rules/

One standing instruction per file, loaded into every session alongside CLAUDE.md. They cover coding standards, git safety, planning and approval gates, agent delegation and team mechanics, prose style, and review discipline. To use one, copy it into your own ~/.claude/rules/.

Where to start

These rules are broadly applicable regardless of setup or stack. The rest of the collection is more tied to my own tooling and workflows.

• take-ownership: exhaust the sources you can reach before asking; verify reviewer and AI findings against the code before acting.
• verify-before-reporting: re-read the artifact before declaring done; every claim needs a source that exists.
• read-whole-files: read load-bearing documents end to end, not sliced or sampled.
• plan-from-current-tree: treat the current tree as the intended state; git history is context, not something to resurrect.
• implement-before-delete: build and verify the replacement before deleting what it supersedes.
• dead-code-closure: deleting a file deletes everything whose last consumer just disappeared.
• generated-files: never hand-edit generated files; fix the source and regenerate.
• secrets: never read tokens or credentials into the conversation, including "checking whether a token exists".
• fix-known-issues: a defect flagged in code the current work touches gets fixed in the same round.
• no-easy-options: recommend the option whose shape is right for the problem, not the cheapest one.
• preserve-user-design: implement the stated design literally; when it hits a wall, present the wall instead of substituting a different design.
• no-invented-scaffolding: don't introduce primitives that were never agreed to; ask instead of inventing.
• engage-framing-first: engage the named approach on its own terms before offering alternatives.
• terse-replies: lead with the answer; save detail for when it's asked for.
• sweep-verification: a corrections sweep gets a verification round before it is called complete.
• no-self-citation: an AI-authored reply is not a settled decision; trace decisions to a human.

skills/

User-invocable workflow skills (/skill-name in a session). Each directory holds a SKILL.md that loads as the operating procedure for that workflow, plus any reference docs it needs:

• jira-ticket-workflow: runs a Jira ticket end to end as a team lead: plan the work, delegate implementation to specialist agents, facilitate cross-review, ship the PR.
• work-the-branch: takes in-progress branch work at any stage, assesses where it stands, and advances it to review-ready.
• pr-review-response: reads every review thread on the current PR, evaluates the feedback critically, implements fixes via agents, and replies to each thread.
• team-dev-workflow: the general-purpose team workflow for any code change, scaled from bug fix to feature.
• epic-orchestrator: dispatches parallel headless Claude Code workers, one per ticket of a Jira Epic, each in an isolated worktree, coordinating through append-only status files.
• session-debrief: an end-of-session candor ritual: what the session is least confident about, and what the user probably doesn't realize.

hooks/

Claude Code hook scripts that run before tool calls. The rules ask the model to behave; these make the load-bearing ones mechanical.

• deny-rule-redirect.js: when a Bash command matches a settings.json deny rule, names the exact rule that fired and the tool to use instead, so the agent self-corrects rather than guessing.
• git-no-verify-block.js: blocks every known form of git hook bypass on commit and push (--no-verify, -n, core.hooksPath overrides via flag or environment).
• yarn-test-root-block.js: blocks yarn test from a monorepo root, where it would run every package's suite; subpackage runs pass.

They register in ~/.claude/settings.json as PreToolUse hooks on Bash:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          { "type": "command", "command": "node ~/.claude/hooks/deny-rule-redirect.js" },
          { "type": "command", "command": "node ~/.claude/hooks/yarn-test-root-block.js" },
          { "type": "command", "command": "node ~/.claude/hooks/git-no-verify-block.js" }
        ]
      }
    ]
  }
}

LSP-first enforcement

My setup also enforces LSP-first code navigation (the grep-banned rule): Grep, symbol-hunting Glob patterns, and shell grep on code symbols are blocked, and Reads of code files are gated behind language-server warmup. The hooks that do this come from claude-code-lsp-enforcement-kit (MIT) and aren't mirrored here; its installer copies the scripts into ~/.claude/hooks/ and merges their settings.json registration.