Claude Code

Anthropic’s terminal-native coding agent. Spec page: /tools/claude-code.

• Strengths: strong multi-file reasoning and coordinated edits; sub-agents keep the main context lean. See context management.
• The biggest multiplier is persistent memory: a directory of corrections read at session start means every fix compounds.
• Cost watchout: token economics dominate. After a “nerfed” model update the same job can take 2-3x the tokens (see the 2,000-hours notes).
• Locked to Claude models, so it does not offer BYOK and model choice.
• Supports MCP, placing it in the modern tool-use ecosystem.

The leverage play is agent leverage and loops: give it success criteria and let it loop. The cost is comprehension debt and skill atrophy, since it is easy to stop reviewing what one-shots. Workflow tip: treat it like a high-leverage junior, with tight specs and a reviewed diff every turn.

Sources: 0001-claude-code-2000h.md, 0007-karpathy-claude-workflow.md, 0008-agentic-coding-trap.md, 0010-self-improving-code-review.md

Edit wiki/claude-code.md; see how we maintain the wiki.