The claim
Rakesh Gohel's post argues that faster models only move the bottleneck: planning, review and approval were sized for human-written code, so the lifecycle itself is what to redesign next. He contrasts two answers. AWS Labs' AI-DLC makes the workflow adapt to each task, so small changes skip stages. Anthropic's AI-native SDLC playbook makes every stage commit an artifact the next stage reads, with humans at the key gates.
The post itself is a framing plus a five-step starter kit: pick one low-risk service, add the AI-DLC rules and check the agent asks questions before coding, write the intent as a repo file and approve the plan first, put conventions in CLAUDE.md and enforce the rest with hooks, and name one senior owner who signs off production. The substance is in the playbook, so most of this summary is about it.
The playbook in one paragraph. Louis Claxton (Anthropic Applied AI), published 2026-08-21, about 40 minutes. When build collapses to hours, plan, review/test and deploy still run at human speed, line-by-line review stops scaling, and governance by committee gets more expensive. The fix is a loop of six non-linear stages (Plan, Design, Build, Test, Deploy, Maintain), each made of "plays" with prerequisites, steps, governance and a leading and lagging metric. The thread through all of it is the committed artifact: intent.md, spec.md, plan.md, the diff and tests, the PR with review findings, the incident record. Each accepted artifact triggers the next stage, and the chain of commits is the audit trail. The end state is closed: a deterministic monitor breaches a control band, Claude diagnoses, and the finding re-enters as a new intent.md.
What is genuinely useful, beyond the product tour.
Every play carries a metric read from git or PR metadata: intent survival rate, spec commits after the first plan commit (requirements rework), merged diff still matching
plan.md, repeat mistakesCLAUDE.mdshould have caught.An explicit advisory-versus-deterministic split: a skill makes a policy likely, a hook makes it close to certain, and a policy that must always hold needs both.
Agent configuration (
CLAUDE.md, skills, hooks) is regression-tested by a CI eval suite of 20 to 50 real tasks, and every incident becomes a permanent eval.Protect the oracle: for bug fixes, commit a failing test first, then block test-file edits during the fix with a hook.
Separation of duties: the agent that wrote the code cannot approve it; branch protection keeps a human code owner.
Maintain stage: detection stays deterministic (rolling mean and sigma with Western Electric rules). At 1 sigma log, at 2 sigma diagnose read-only, at 3 sigma propose only through a PR or a pre-approved runbook.
A sidebar on legacy systems: for each artifact, name one source of truth (repo, or Jira and similar) and at minimum link record IDs and commit SHAs both ways.
The artifact chain
An accepted artifact fires the next stage; humans sit at the gates rather than starting each stage. Only Maintain runs without a person in the invocation path.
Playbook vs AI-DLC
They are complements more than rivals: AI-DLC is an engine you install, the playbook is an operating model you assemble from Claude Code primitives. AI-DLC details come from its 2.x repo and spec; playbook details from the article.
| Dimension | Anthropic AI-native SDLC playbook | AWS Labs AI-DLC |
|---|---|---|
| What it is | Guidance: six stages of plays, each with steps, governance and metrics | Open-source rules plus a deterministic workflow engine for coding agents (MIT-0) |
| Who routes the work | Accepted artifacts trigger the next stage (merge of intent.md, approval of spec.md, merged PR) | The engine routes phases and stages, and skips stages for small tasks |
| Unit of record | Committed Markdown per change: intent, spec, plan, then diff, PR and incident | Per-intent records, an audit log, kept rules and a code-derived knowledge base |
| Human gates | Product owner on intent and spec, engineer on plan, code owner on PR, release manager at production | Explicit stage gates with recorded approvals bound to a source fingerprint |
| Enforcement | Skills (advisory) backed by hooks and managed settings (deterministic) | Engine routing plus sensors and post-conditions on stage outputs |
| Learning loop | A mistake made twice goes into CLAUDE.md; every incident becomes an eval | A human correction becomes a consented, dated rule after a conflict check |
| Closing the loop | Monitors and scheduled scans write new intent.md files without a human starting them | Not a focus; work starts from a human request |
| Measurement | A leading and lagging metric per play, mostly from git and PR metadata | Approval and audit records; less on outcome metrics |
| Portability | Written for Claude Code, Claude Tag and Claude Security, but the pattern is tool-neutral | One core across several harnesses |
The post's starter kit mixes them sensibly: AI-DLC's rules for structure on day one, the playbook's artifacts, CLAUDE.md and hooks for the parts you will want to own.
Tradeoffs and failure modes
Vendor playbook. Every play lands on an Anthropic product (Claude Design, Code Review, Claude Security on Mythos 5, Claude Tag). The pattern ports; the infrastructure lists do not. Read it as a reference architecture, not neutral research.
No evidence of outcomes. Each play names metrics but the article reports none: no before and after, no customer numbers. The "weeks to hours" claims are expectations.
Artifact sprawl and drift. Four Markdown files per change plus
CLAUDE.md, skills andREVIEW.mdis a lot of prose to keep true. The playbook only half-addresses drift (updateplan.mdin the same commit, "consider a hook");intent.mdandspec.mdcan quietly go stale after build starts.Rationale stays per change. Like AI-DLC, nothing consolidates decisions across changes into a system-level view. Six months in, the why is scattered across hundreds of spec files.
Review moves, it does not vanish. Humans review intent, spec, plan and PR. With parallel sessions the queue reappears at the plan and PR gates; the playbook's own ceiling is "as many streams as one person can review properly".
Model-on-model review. Agentic review passes and an adversarial reviewer share priors with the writer. The deterministic layers (tests the agent cannot edit, hooks, branch protection) carry the real guarantee.
Closed-loop risk. At 3 sigma the agent may quarantine a flaky test or open a revert. Quarantining tests autonomously can hide real regressions, and band tuning by dismissal can teach the monitor to ignore a slow failure.
One-size ceremony. The playbook's stages are non-linear, but it has no rule for when a one-line fix skips intent and spec. AI-DLC's adaptive depth is the stronger idea here.
Skills are advisory. The playbook says so itself, which is to its credit: teams that stop at skills get likelihood, not control.
Review checklist
For a team adopting either approach on one service:
Is there one committed intent artifact per change, and does the product owner's accept or reject show in git?
Does each stage start by reading the previous stage's artifact, not a chat summary?
Is there a rule for which changes skip intent and spec (size, risk class, blast radius)?
Is
plan.mdkept in sync with the diff by a hook or a review pass, not by habit?Which policies must always hold, and is each backed by a hook or managed setting rather than only a skill?
Can the agent edit the tests that judge its fix? It should not.
Does a change to
CLAUDE.md, skills or hooks run an eval suite before merge?Is there a human code owner the writing agent cannot replace, enforced by branch protection?
Can the agent reach production? It should stop at a gate a named person opens.
For any autonomous loop: is detection deterministic, is each tier's power written in config, and is rollback rehearsed?
Which system is the source of truth for each artifact, and are record IDs and commit SHAs linked both ways?
Are the per-play metrics actually collected, so the team can tell whether the redesign helped?
Related entries
R017 (Elephant-Goldfish Model): another spec-first delivery workflow; its fresh-context reviewers are a stronger form of the playbook's review passes.
R020 (the model proposes, the harness enforces): the playbook's skill-versus-hook split and its "agent cannot edit the test" rule are R020's thesis applied.
R052 (Alibaba Open Code Review): the PR review stage built as its own system.
R015 (workflows over agents): artifacts that trigger the next stage are a workflow, not an autonomous agent.
R016 (human verification of AI output): the human role left at each gate is auditor, which is what the playbook designs for.
R031 (multi-agent incident response): the Maintain stage's closed loop, from the incident side.
R036 (Claude 101 harness tips): the practical
CLAUDE.md, skills and hooks setup the Build stage assumes.
Sources
Rakesh Gohel, LinkedIn post, 2026-11-12.
The AI-native SDLC playbook, Louis Claxton, Anthropic, 2026-08-21.