Fable Brain

Claude Code Platform

5 min read

Claude Code Platform

Claude Code is Anthropic’s developer harness, reachable as a CLI, a VS Code extension, a JetBrains plugin, Desktop, and Web, documented canonically at code.claude.com/docs, and it serves Claude Fable 5 to Pro, Max, Team, and Enterprise users.

What it is

The agentic coding surface built around a terminal-native agent loop. Its canonical documentation moved: docs.claude.com Claude Code pages 301-redirect to code.claude.com/docs, and the former engineering post at anthropic.com/engineering/claude-code-best-practices 308-redirects to code.claude.com/docs/en/best-practices, making 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 same engine is exposed programmatically as the Claude Agent SDK.

How it works

  • Surfaces. The VS Code extension (also installable in Cursor and other forks) is the recommended IDE integration: “This is the recommended way to use Claude Code in VS Code.” (https://code.claude.com/docs/en/vs-code). 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.
  • Memory. Two systems: CLAUDE.md files loading across managed, user, project, and local scopes with sizing guidance to “target under 200 lines per CLAUDE.md file”, plus auto memory (default-on since v2.1.59), where Claude saves per-repository learnings under ~/.claude/projects//memory/ and the first 200 lines or 25KB of MEMORY.md load into every session.
  • Control plane. Six permission modes (default Manual, acceptEdits, plan, auto, dontAsk, bypassPermissions) cycled with Shift+Tab; settings merge across a documented hierarchy; hooks execute deterministically at lifecycle events; subagents run in separate context windows; custom slash commands and skills are one system following the Agent Skills open standard.
  • Automation. Non-interactive mode (claude -p) is the official path for CI and pipelines with text, json, and stream-json output, and “—bare is the recommended mode for scripted and SDK calls” (https://code.claude.com/docs/en/headless).
  • Model availability. Claude Fable 5 is deployed in Claude Code for Pro, Max, Team, and Enterprise users since the June 9, 2026 launch, with the June 12 to July 1 export-control interruption.
  • Context split by load pattern. Official guidance puts broadly applicable, every-session facts in CLAUDE.md and moves multi-step procedures into skills, since “a skill’s body loads only when it’s used” (https://code.claude.com/docs/en/skills); CLAUDE.md also supports @path imports up to four hops.

Key insight

The platform’s design theme is determinism around a probabilistic core: hooks execute at lifecycle events regardless of what the model decides (“Unlike CLAUDE.md instructions which are advisory, hooks are deterministic and guarantee the action happens.”, https://code.claude.com/docs/en/hooks), permission rules merge across scopes, and exit code 2 from a hook blocks the action while stderr feeds back to Claude.

Best practice

  • Read platform behavior from code.claude.com/docs, not from older anthropic.com or docs.claude.com URLs; both redirect. EVIDENCE-BASED
  • Pick the surface by workflow, IDE extension for inline review, CLI for terminal-native work, headless -p for automation; configuration is shared underneath. EVIDENCE-BASED
  • Restrict bypassPermissions to isolated environments; the docs say to use it only in containers, VMs, or dev containers, and it refuses to start as root. EVIDENCE-BASED
  • Treat auto mode (research preview, v2.1.83+) as a monitored middle ground: “A separate classifier model reviews actions before they run” (https://code.claude.com/docs/en/permission-modes), and it pauses back to prompting after repeated blocks. EVIDENCE-BASED
  • Constrain through systems, hooks and permission rules, rather than prompt prose, per converging official and practitioner guidance. PRACTITIONER
  • Review auto memory with /memory periodically; the first 200 lines or 25KB of MEMORY.md enter every session, so stale self-written notes compound. EVIDENCE-BASED
  • For CI, pre-approve tools with —allowedTools permission-rule syntax and use —json-schema when downstream steps parse output. EVIDENCE-BASED

Pitfalls

  • Citing the old best-practices blog URL as a source; it is a redirect, and content ownership now sits in the docs.
  • Assuming AGENTS.md works: “Claude Code reads CLAUDE.md, not AGENTS.md.” (https://code.claude.com/docs/en/memory); the official pattern is a CLAUDE.md that imports it.
  • Forgetting auto memory exists; Claude writes its own per-repo notes, inspectable via /memory and disabled with autoMemoryEnabled false.
  • Expecting the /agents creation wizard; as of v2.1.198 it is removed in favor of asking Claude or editing subagent files directly.
  • Applying claude.ai export rules to this harness; the system prompt capture documents the consumer surface, and this platform’s contract lives in its docs.

Sources

Next actions

  • Record the installed Claude Code version and which documented features (auto mode, auto memory) it carries.
  • Map this vault’s own hook and settings usage against the documented hierarchy.
  • Cross-check Export Omits Claude Code Harness once the corpus chapter notes are complete.
  • Inventory this project’s auto memory contents via /memory and prune stale self-written notes.

Graph View

Backlinks