Fable Brain

Claude Code

6 min read

Claude Code

Claude Code is Anthropic’s agentic coding harness, shipped as a CLI with VS Code, JetBrains, Desktop, and Web surfaces, and its canonical documentation now lives at code.claude.com/docs.

What it is

Claude Code is the agentic coding tool documented at code.claude.com/docs. The old docs.claude.com Claude Code pages 301-redirect there, and the former engineering post at anthropic.com/engineering/claude-code-best-practices 308-redirects to code.claude.com/docs/en/best-practices, which makes the docs page the current official best-practices guide. The docs cover memory, permission modes, settings, hooks, subagents, skills, non-interactive mode, and per-surface guides for CLI, VS Code, JetBrains, Desktop, and Web. The source repository is github.com/anthropics/claude-code.

The VS Code extension (also installable in Cursor and other forks) is the recommended IDE integration. It provides inline diffs, @-mentions with line ranges, plan review, and checkpoints, bundles its own CLI copy, and shares ~/.claude/settings.json, conversation history, CLAUDE.md, and MCP config with the terminal CLI. The JetBrains plugin runs the CLI in the IDE terminal with the same Shift+Tab mode switching.

The claim packs verify Claude Code mechanics but say nothing Fable 5 specific about its deployment in this harness. The system prompt export describes claude.ai, not Claude Code. See Export Omits Claude Code Harness and Which Export Rules Bind Claude Code before assuming export rules apply here.

How it works

  • Project memory. CLAUDE.md files load at defined scopes in order, managed policy (for example /etc/claude-code/CLAUDE.md on Linux), user ~/.claude/CLAUDE.md, project ./CLAUDE.md or ./.claude/CLAUDE.md, and gitignored ./CLAUDE.local.md. Parent-directory files load in full at launch, subdirectory files load on demand. @path imports recurse up to four hops. “Claude Code reads CLAUDE.md, not AGENTS.md.” (https://code.claude.com/docs/en/memory); repos using AGENTS.md import it from CLAUDE.md or symlink it.
  • Auto memory. A second memory system, on by default since v2.1.59, stores “notes Claude writes itself based on your corrections and preferences” (https://code.claude.com/docs/en/memory) under ~/.claude/projects//memory/. The first 200 lines or 25KB of MEMORY.md load every session; inspect via /memory, disable with autoMemoryEnabled false.
  • Permission modes. Six modes, default (Manual), acceptEdits, plan, auto, dontAsk, bypassPermissions, cycled with Shift+Tab or set via —permission-mode and the defaultMode setting. Plan mode is read-only research; approving the plan switches into the chosen execution mode, and Ctrl+G opens the plan in a text editor.
  • Auto mode. A research preview (v2.1.83+) where “A separate classifier model reviews actions before they run” (https://code.claude.com/docs/en/permission-modes), blocking scope escalation, unknown infrastructure, and hostile-content-driven actions. It pauses back to prompting after 3 consecutive or 20 total blocks; in -p runs repeated blocks abort the session.
  • Settings hierarchy. Managed org settings override command-line arguments, then .claude/settings.local.json, then project .claude/settings.json, then user ~/.claude/settings.json. Permission rules merge across scopes rather than override, and most edits hot-reload.
  • Hooks. Deterministic shell commands at lifecycle events (SessionStart, SessionEnd, UserPromptSubmit, Stop, PreToolUse, PostToolUse, PermissionRequest, compaction events), configured in settings JSON with matchers. Exit code 2 blocks the action and feeds stderr back to Claude.
  • Subagents. Markdown files with YAML frontmatter (name, description, tools, model) in .claude/agents/ or ~/.claude/agents/. “Each subagent runs in its own context window with a custom system prompt” (https://code.claude.com/docs/en/sub-agents). Built-in Explore, Plan, and general-purpose subagents are delegated to automatically; the /agents wizard is removed as of v2.1.198.
  • Skills and commands. “Custom commands have been merged into skills.” (https://code.claude.com/docs/en/skills). A .claude/commands/deploy.md file and a .claude/skills/deploy/SKILL.md both create /deploy; skills add supporting files, disable-model-invocation control, $ARGUMENTS, and dynamic context injection, and follow the Agent Skills open standard.
  • Headless. claude -p is the official CI and scripting path, with text, json, and stream-json output, —json-schema, —allowedTools pre-approval, —continue/—resume, and —bare to skip auto-discovery of hooks, skills, plugins, MCP servers, and CLAUDE.md for reproducible runs.

Best practice

  • Give Claude a verifiable check, tests, a build exit code, a linter, or a screenshot comparison, so it iterates until the check passes instead of using you as the verification loop. EVIDENCE-BASED
  • Follow explore, plan, implement, commit; use plan mode to research first, but “If you could describe the diff in one sentence, skip the plan.” (https://code.claude.com/docs/en/best-practices). EVIDENCE-BASED
  • Keep CLAUDE.md ruthlessly pruned to what Claude cannot infer from code, target under 200 lines per file, and run /init for a starter. EVIDENCE-BASED
  • Enforce mandatory rules with hooks and permission rules, not prose; “Unlike CLAUDE.md instructions which are advisory, hooks are deterministic and guarantee the action happens.” (https://code.claude.com/docs/en/hooks). EVIDENCE-BASED
  • Split persistent context by load pattern, every-session facts in CLAUDE.md, multi-step procedures in skills, because “a skill’s body loads only when it’s used” (https://code.claude.com/docs/en/skills). EVIDENCE-BASED
  • Run /clear between unrelated tasks, and after roughly two failed corrections start a fresh session with a better prompt. EVIDENCE-BASED
  • Restrict bypassPermissions to isolated environments such as containers, VMs, or dev containers; it refuses to start as root. EVIDENCE-BASED
  • Make policy hooks exit with code 2; “Exit 1 does not block anything.” (https://blakecrosley.com/blog/claude-code-hooks-explained). PRACTITIONER
  • Prototype fan-out batch runs on 2 to 3 files before scaling to the full set. PRACTITIONER

Pitfalls

  • “Bloated CLAUDE.md files cause Claude to ignore your actual instructions!” (https://code.claude.com/docs/en/best-practices). Overlong memory files reduce adherence.
  • Expecting AGENTS.md to be read natively; it is not, import or symlink it.
  • Writing policy hooks that exit 1 and assuming they block; only exit 2 blocks.
  • Relying on auto mode in unattended -p runs; repeated classifier blocks abort the session.
  • Assuming a lower-scope settings file wins; managed org settings sit above everything, including CLI flags.
  • Scripting against the /agents creation wizard; it is removed in v2.1.198 in favor of editing agent files directly.

Sources

Next actions

  • Capture Fable 5 specific Claude Code behavior (model IDs, defaults) from official release notes when available.
  • Verify the v2.1.x feature gates (auto memory v2.1.59, auto mode v2.1.83, wizard removal v2.1.198) against the CHANGELOG in github.com/anthropics/claude-code.
  • Cross-check hook lifecycle event names against a live claude install on Aurora.

Graph View

Backlinks