Fable Brain

Claude Code Quickstart for Fable 5

5 min read

Claude Code Quickstart for Fable 5

Install Claude Code from the official per-surface guides, seed a short CLAUDE.md with /init, start work in plan mode, tighten permissions in settings, then run the explore, plan, implement, commit loop with a verifiable check.

What it is

  • A zero-to-productive path for running Claude Fable 5 inside Claude Code, built only from claims verified on 2026-07-07.
  • Fable 5 is deployed in claude.ai and Claude Code for Pro, Max, Team, and Enterprise users (https://platform.claude.com/docs/en/about-claude/models/introducing-claude-fable-5-and-claude-mythos-5).
  • The canonical documentation lives at code.claude.com/docs: docs.claude.com Claude Code pages 301-redirect there, and the former engineering best-practices post 308-redirects to code.claude.com/docs/en/best-practices.

How it works

  1. Install and pick a surface. The docs carry per-surface guides for CLI, VS Code, JetBrains, Desktop, and Web (https://code.claude.com/docs/en/overview). For IDE work the VS Code extension is “the recommended way to use Claude Code in VS Code.”; it bundles its own CLI copy while sharing ~/.claude/settings.json, history, CLAUDE.md, and MCP config with the terminal.
  2. Seed project memory. Run /init to generate a starter CLAUDE.md; include only what Claude cannot infer from code (build commands, divergent style rules, repo etiquette, gotchas) and target under 200 lines per file. Scopes load in order: managed policy, user ~/.claude/CLAUDE.md, project ./CLAUDE.md or ./.claude/CLAUDE.md, and gitignored ./CLAUDE.local.md. @path imports recurse up to four hops; “Claude Code reads CLAUDE.md, not AGENTS.md.” (https://code.claude.com/docs/en/memory).
  3. Know the second memory. Auto memory is on by default since v2.1.59: Claude saves per-repository learnings under ~/.claude/projects//memory/, the first 200 lines or 25KB of MEMORY.md load each session, /memory inspects it, and autoMemoryEnabled false disables it.
  4. Learn the six permission modes: default (Manual), acceptEdits, plan, auto, dontAsk, bypassPermissions, cycled with Shift+Tab or set via —permission-mode and defaultMode. Plan mode is read-only research: “Plan mode tells Claude to research and propose changes without making them.”, entered via Shift+Tab, a /plan prefix, or claude —permission-mode plan, with Ctrl+G opening the plan in an editor (https://code.claude.com/docs/en/permission-modes).
  5. Configure settings. Priority order: managed org settings, then command-line arguments, then .claude/settings.local.json, then project .claude/settings.json, then user ~/.claude/settings.json; permission rules merge across scopes and most edits hot-reload (https://code.claude.com/docs/en/settings).
  6. Run the first agentic loop: explore and plan in plan mode, approve into an execution mode, implement, commit. Give the model a check it can run, “tests, a build, a screenshot to compare”, so it iterates until the check passes instead of using you as the verification loop.

Best practice

  • Skip planning overhead on small fixes: “If you could describe the diff in one sentence, skip the plan.” (https://code.claude.com/docs/en/best-practices). EVIDENCE-BASED
  • Reserve bypassPermissions for isolated environments like containers or VMs; it refuses to start as root (https://code.claude.com/docs/en/permission-modes). EVIDENCE-BASED
  • Try auto mode (research preview, v2.1.83+) for prompt fatigue: a separate classifier model reviews actions, blocking scope escalation and hostile-content-driven actions, pausing after 3 consecutive or 20 total blocks. EVIDENCE-BASED
  • Enforce must-happen rules with hooks, not CLAUDE.md prose; hooks run deterministically at lifecycle events and exit code 2 blocks the action (https://code.claude.com/docs/en/hooks). EVIDENCE-BASED
  • Put every-session facts in CLAUDE.md and multi-step procedures in skills, since “a skill’s body loads only when it’s used” (https://code.claude.com/docs/en/skills). EVIDENCE-BASED
  • Isolate research and verification in subagents; each runs in its own context window with a custom system prompt (https://code.claude.com/docs/en/sub-agents). EVIDENCE-BASED
  • For CI and scripts use claude -p with —allowedTools, json output, and the —bare flag, “the recommended mode for scripted and SDK calls” (https://code.claude.com/docs/en/headless). EVIDENCE-BASED
  • Prototype fan-out batch runs on 2 or 3 files before scaling to the whole repo. PRACTITIONER

Pitfalls

  • Bloated CLAUDE.md files cause Claude to ignore your actual instructions; ruthless pruning is official guidance, with IMPORTANT or YOU MUST emphasis reserved for critical rules.
  • Hook scripts that exit 1 expecting a block; “Exit 1 does not block anything.” Policy hooks must exit 2 (https://blakecrosley.com/blog/claude-code-hooks-explained).
  • Correcting the same failure repeatedly in one session; official guidance says a clean session with a better prompt almost always outperforms a long session.
  • Expecting AGENTS.md to load; the official pattern is a CLAUDE.md that imports it or a symlink.
  • Looking for the /agents creation wizard; as of v2.1.198 it is removed in favor of asking Claude or editing agent files directly.
  • In non-interactive auto-mode runs, repeated classifier blocks abort the session; plan permissions ahead for unattended work.

Sources

Next actions

  • Verify the quickstart end-to-end on a fresh machine and note any drift from the docs.
  • Add a worked example of a PreToolUse hook with exit 2 blocking.
  • Track auto mode’s exit from research preview status.

Graph View

Backlinks