Anthropic’s engineering team published how they actually run Claude Code on real codebases: context files, permission setup, test-driven loops, multi-agent workflows and the explore-plan-code-review rhythm. It reads like a field manual, not a brochure - and most of it transfers to any coding agent you use.
What you get
- The context-file playbook. What to put in
AGENTS.md/CLAUDE.md, what to leave out, and how the file compounds over time. - Delegation patterns. When to pair interactively, when to hand over a task end-to-end, and how to structure long sessions so they stay verifiable.
- Verification loops. Using tests as the agent’s feedback signal - the single habit that most improves output quality.
How to use it
Read it once, then adopt in this order: (1) write your context file, (2) make the agent run tests after changes, (3) practice the explore → plan → code → review rhythm on one real task. Each step pays before the next one costs anything.
Good to know
It is vendor-written, so treat tool-specific claims accordingly - but the workflow content is the closest thing to a consensus reference for agentic coding. Pair it with our AGENTS.md news analysis for the portable version of the same ideas.