Anthropic publishes a playbook for running the SDLC around coding agents

Anthropic's AI-native SDLC playbook proposes chaining versioned intent.md, spec.md and plan files so each approved artifact triggers the next stage.

Anthropic's AI-native SDLC playbook (published August 21, 2026) argues that once agents write most of the code, the bottleneck moves to planning, review, testing and deploy, which "still run at human speed." Its answer is a working pattern a small team can copy: each stage ends by committing a markdown or code artifact, and the next stage begins by reading it.

The chain Anthropic describes runs intent.md → spec.md → a plan built in Claude Code plan mode → the diff and its tests → the PR with review findings → the incident record. An accepted intent.md triggers the spec pass, an approved spec triggers plan mode, a merged PR triggers the pipeline, and a production alert writes the next intent.md. The commit history then works as the audit trail. Organization policy for brand, security and compliance is encoded as skills that Claude applies while writing the spec. The playbook includes a sample intent file and a spec prompt you can adapt. Anthropic recommends prompting each step by hand first and automating the triggers later.

The playbook keeps humans at the gates. A product owner corrects and approves the intent and the spec, and higher-risk work goes to a technical lead.

The limits matter. This is a vendor document built around Claude Code, Cowork and Claude Design (beta), and it reflects Anthropic's Applied AI team's practice, not independent results. A summary by @bibryam on X also credits it with hooks for enforcement and continuous evals gating deploys. We haven't verified those details against the later sections. There's also no outside data yet on review load or PR quality when teams adopt this approach.

Sources