claude.dev-first engineering playbook
Anthropic's teams publish how they build software with agents on claude.dev. This skill carries that guidance, digested post by post. Current through 2026-10-15 (16 posts).
How to answer
- Find the question's topic in the table below and read that reference file before answering. Open a post's digest in
posts/when you need its details or exact words. - Shape the answer like this:
- What Anthropic's teams do: the guidance that applies, each point linked to its post as
https://claude.dev/blog/<slug>/. - Beyond claude.dev: anything the posts don't cover, answered from general practice and labeled as such. Leave this part out when the posts cover the whole question.
- What Anthropic's teams do: the guidance that applies, each point linked to its post as
- Quote only from the digests in
posts/, never from memory. - A digest's "Perishable details" (model names, prices, effort defaults) hold only for the model or version they name. Check what's current before repeating one.
- If the user asks for the newest guidance, or the question concerns something after the date above, read https://claude.dev/rss.xml and any post newer than that date before answering.
Topics
| The question is about | Read |
|---|---|
| Code review, approvals, merging, flags, rollout, rollback, who owns a change | references/shipping.md |
| Tests, evals, golden cases, graders, benchmarks, CI gates, agents checking their own work | references/verification.md |
| Handing work to agents: task specs, CLAUDE.md, effort, model choice, local or cloud sessions, subagents, skills, output formats | references/delegation.md |
| Agents running unattended or on a schedule: state between runs, credentials, budgets, guardrails | references/automation.md |
| Building agents: tools, context, prompt caching, workflows | references/agent-design.md |
| Cost per task, human time per change, reverts, incidents | references/measurement.md |
Core rules
- Name the finish line and the check that proves it, then let it cook. (Opus 5.5 playbook, Cloud sessions)
- No claim without evidence. Agents run real checks and show the output before a person looks, and a different agent does the verifying. (Code review, Dynamic workflows)
- Ask which check should catch a problem, not whether a human read the code. Human approvals scale with risk, the change sets the tier, and an agent's approval never counts as a person's. (Code review)
- Judge the system by human time per change, with reverts and incidents as the guardrail. (Code review)
- Golden cases come from reality, and two experts would agree on each verdict. Turn incidents into eval cases, and read graded transcripts before trusting a grader. (Code review, Eval design)
- Trust only gains that beat the noise on cases the tuner never saw. (Eval design)
- Give agents a number to beat that is proven to track the real outcome, and ratchet every win in CI. (claude.ai 3x faster)
- Make mistakes cheap to undo: a merge queue, short-lived flags, staged rollout, fast rollback, and a named owner for every change. (Code review, claude.ai 3x faster)
- Unattended agents fail loudly: a failed read is never a quiet day, and nothing counts as done until the destination confirms it. (Agent automations)
- Keep context lean and let the agent find the rest; delete rules as models improve. (Context engineering, Seeing like an agent)
- Pay for finished tasks, not tokens: a check is cheaper than more effort, and a retry costs more than any saving. (Task cost)