Claude Code commands, flags, and config in one place

Headless and CI: print mode flags

Print mode is what turns Claude Code from a chat tool into a pipeline step. These flags are the ones that make an unattended run safe and parseable.

A complete, safe CI invocation looks like this - read-only tools, a turn ceiling, a budget ceiling, and machine-readable output:

claude -p "Review the diff against main and list correctness bugs only." \
  --output-format json \
  --allowedTools "Read" "Grep" "Bash(git diff *)" "Bash(git log *)" \
  --max-turns 6 \
  --max-budget-usd 2.00 \
  --no-session-persistence

Gotcha: --max-turns, --max-budget-usd, and --no-session-persistence are print-mode only. Passing them to an interactive session does nothing, which is easy to miss because nothing errors.

Model selection, effort, and context

Which model runs by default depends on your plan, so set it explicitly when it matters. As of August 2026 the aliases map to Anthropic's current line-up: Fable 5 at the top, Opus 5 for complex agentic coding, Sonnet 5 for the speed and intelligence balance, and Haiku for cheap mechanical work.

Gotcha: the default model is plan-dependent and has changed more than once during 2026. Run /status to see what you are actually talking to before you draw conclusions about quality.

CLAUDE.md memory and project rules

Memory files load at the start of every session, in scope order from broadest to most specific, and all discovered files are concatenated rather than overriding each other.

Claude Code reads CLAUDE.md, not AGENTS.md. If your repo already has AGENTS.md for other agents, bridge them with a one-line import so nothing is duplicated:

@AGENTS.md

## Claude Code

Use plan mode for changes under `src/billing/`.
Never run the deploy script; ask instead.

Gotcha: target under 200 lines per file. Anthropic's own guidance is that longer files consume more context and reduce adherence - so a bloated CLAUDE.md costs you money and obedience at the same time. Move procedures into skills and file-specific rules into .claude/rules/.

Hooks: the enforcement layer

A hook binds a command, an HTTP call, an MCP tool, a prompt, or a sub-agent check to a lifecycle event. Unlike a CLAUDE.md rule, it runs regardless of what the model decided. There are more than thirty events; these are the ones people actually use.

Hooks live in settings.json under a hooks key. This one formats every file the agent writes:

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Write|Edit",
        "hooks": [
          {
            "type": "command",
            "command": "jq -r '.tool_input.file_path' | xargs npx prettier --write",
            "timeout": 60
          }
        ]
      }
    ]
  }
}

Gotcha: hook input arrives as JSON on stdin, so a naive command that expects a file path as an argument will silently do nothing. Parse the payload. The if field takes a permission-rule pattern such as Edit(*.ts) if you want to narrow further than the matcher.