Code Is No Longer the Bottleneck: An AI-Native SDLC Playbook in Six Stages

Forum topic · QianXun · 2026-08-28

Summary

A full Chinese translation and stage-by-stage breakdown of Anthropic's Applied AI team playbook on the AI-native software development lifecycle (SDLC), authored by Louis Claxton. The core thesis: agentic coding tools like Claude Code have collapsed the cost of writing code, so the bottleneck has shifted to the human-paced stages around Build—planning, review/testing, and deployment. The playbook restructures the traditional six SDLC stages (Plan, Design, Build, Test, Deploy, Maintain) around versioned artifacts that chain together as an audit trail: intent.md, spec.md, plan.md, diffs with tests, PRs with review verdicts, and incident records. Each of the six Plays explains what changed, how to start, concrete steps, governance considerations, and success metrics. Key practices include capturing intent directly from proposers as intent.md, collapsing requirements and design into a single Claude session, using Claude Code plan mode before any code is written, encoding institutional knowledge in CLAUDE.md and skills, enforcing non-negotiable policies with hooks, running parallel sessions in git worktrees, applying layered agentic review at deploy time, and feeding broken controls from production back into new intent.md files to close the loop.

Key points

This post is a full annotated translation of an Anthropic Applied AI team playbook (author: Louis Claxton) on rebuilding the software development lifecycle (SDLC) for agentic AI with Claude and Claude Code.

Core thesis: Organizations now write code at previously unimaginable speeds, but approval gates, reviews, handoffs, and policies still run at human pace. When Build collapses to hours, the bottleneck moves to the stages on either side—Plan, review/Test, and Deploy—and line-by-line human code review no longer scales when diffs are mostly agent-produced.
AI-native SDLC structure: The traditional six stages (Plan, Design, Build, Test, Deploy, Maintain) become a loop with AI embedded at every step, automated handoffs, and automatic triggering of the next Play.
The artifact chain as audit trail: Every stage ends by committing a deliverable to version control, and the next stage begins by reading it:

intent.md → spec.md → plan.md → diff + tests → PR + review verdict → incident record

Humans remain accountable for judgment decisions; their attention moves to the gates the review agents flag.

Stage highlights

01 Plan: Proposers brainstorm directly with Claude to produce an intent.md in their own words (problem, outcome, affected users, constraints, open questions), stored in a shared versioned location (e.g., an intent/ directory in the product repo). A product owner reviews before acceptance. Metric: weeks of elicitation compress to hours; survival rate of intents tracks quality.
02 Design: Requirements and design collapse into one session. Claude, guided by organization skills encoding brand, security, compliance, and UX policy, turns intent.md into a spec.md with concerns flagged. Policy is applied at authoring time, not discovered weeks later at review. A beta flow uses Claude Design for frontend mockups exported to Claude Code.
03 Build: Engineers start every change in Claude Code plan mode; nothing is implemented without an accepted plan committed as plan.md. Institutional knowledge lives in a versioned, one-page CLAUDE.md (rule: if Claude makes the same mistake twice, fix it in CLAUDE.md). Skills (.claude/skills/<name>/SKILL.md) encode knowledge that must be applied consistently—though skills are persuasive controls only, so non-negotiable policies are backed by deterministic hooks that block protected paths, run formatters, and keep credentials out of diffs. Engineers run parallel sessions in separate git worktrees and delegate recurring work to sub-agents (.claude/agents/). A sidebar covers coexistence with legacy systems (Jira, ServiceNow, Figma) via a designated single source of truth or bidirectional links with commit SHAs.
04 Test: QA gates at stage boundaries are replaced by continuous evals woven into implementation; feedback loops let sessions self-verify.
05 Deploy: Layered agentic review with human review reserved for regulated and critical code. Governance executes when AI acts—hooks serve as approval gates—and CI/CD pipelines are triggered by merged PRs. The final diff is checked against the accepted plan.md.
06 Maintain: Agents monitor production deployments; any broken control boundary is diagnosed and written back as a new intent.md, closing the loop.

How to adopt

Plays are modular and list their prerequisites; start from any Play with no incoming arrows (no prerequisites)—typically Plan or the CLAUDE.md/Skills Plays in Build.
Adoption begins with manually prompting each step; the end state is a loop where each accepted deliverable fires the next gate, with human attention focused on reviewing flagged items.
Each Play defines leading metrics (e.g., wall-clock time between artifact commits) and lagging metrics (e.g., rework rounds per change, rate of first-pass merges), read mostly from git history and PR metadata.

The article's throughline: intent, specs, plans, diffs, and review verdicts together form the audit trail—the loop can run autonomously while human judgment sits above it.