ai-sdlc-claude

The playbook, operationalised

Source: Anthropic Academy, The AI-Native SDLC Playbook. This page maps each play to what the sdlc plugin gives you: the command, the artifact, the gate, the evidence, and the metric. Stages are a loop, not a line.

flowchart LR
  I[idea · ticket · incident] --> P1
  subgraph loop
    P1[1 Plan<br/>/sdlc:intent → intent.md] -->|PO accepts| P2[2 Design<br/>/sdlc:spec → spec.md]
    P2 -->|PO accepts| P3[3 Build<br/>/sdlc:plan → plan.md<br/>/sdlc:build → diff]
    P3 --> P4[4 Test<br/>/sdlc:verify · evals]
    P4 --> P5[5 Deploy<br/>/sdlc:review · /sdlc:babysit<br/>/sdlc:release]
    P5 -->|release manager| P6[6 Maintain<br/>control bands · /sdlc:incident]
    P6 -->|new intent.md| P1
  end

Stage	Play	Command(s)	Artifact committed	Gate (human)	Evidence	Leading / lagging metric
1 Plan	Capture intent	/sdlc:intent (Claude Code, claude.ai or Cowork)	sdlc/changes/<id>/intent.md	product owner merges with status: accepted	git history of the intent (author, timestamp, revisions)	hours from first conversation to committed intent / survival rate
2 Design	Requirements & design	/sdlc:spec or CI sdlc-spec-on-intent	spec.md with flagged concerns, skills_applied	product owner (tech lead if high risk)	spec, prompt, skill versions in git	intent→spec hours / spec edits after plan
3 Build	Plan mode	/sdlc:plan	plan.md	engineer accepts in plan mode (tech lead if high risk)	plan + revisions, who accepted	first-pass merge share / rework cycles, plan-diff drift
3 Build	CLAUDE.md	/init, then one page	CLAUDE.md	code owners in PR	git history	repeated mistakes / new joiner time-to-first-PR
3 Build	Skills	skills/policies/* templates → .claude/skills/	SKILL.md per policy	policy owner signs skill PR	skill invocations in traces	policy→skill merge time / policy findings → 0
3 Build	Parallel sessions & subagents	claude --worktree, agents/*	.claude/agents/*.md	—	OTel export attribution	concurrent sessions while review holds / merges per engineer vs rework
3 Build	Hooks as guardrails	plugin hooks: protected-paths, secrets-guard, format-on-edit	.sdlc/gate.log	—	hook decisions	—
4 Test	Feedback loop	/sdlc:verify, --fix, --ui	test/build/lint output in the PR; .sdlc/fix-lock	code owner	toolchain output in transcript + PR	first-pass CI success / review time per PR, change failure rate
4 Test	Continuous evals	/sdlc:eval-add, CI sdlc-evals	evals/cases/*.json, results	config owner approves config PR	eval run logs	pass rate over time, incident→eval time / regressions caught in CI vs prod
5 Deploy	AI in PR review	CI sdlc-review, /sdlc:review, @claude, /sdlc:babysit	PR findings + tally	code owner approval via branch protection	PR history	time to first review, comments resolved without a human / defects caught vs escaped
5 Deploy	Hooks as approval gates	production-gate, managed settings	.sdlc/gate.log, CI environment approvals	release manager	timestamped allow/block	wait per gate / violations reaching production
5 Deploy	CI/CD integration	sdlc-ci-triage, sdlc-deploy, /sdlc:release, /sdlc:rollback	triage comments, release PR, tags	environment protection	pipeline logs under the agent’s identity	failures triaged without paging / DORA
6 Maintain	Closing the loop	monitoring/detect.py, CI sdlc-control-band, /sdlc:incident	sdlc/incidents/<id>/intent.md, lessons.md	service owner triage	detection log with breach timestamp + tier	breach→intent time / findings that become merged fixes, repeat incidents
6 Maintain	Claude on call	RingCentral connector + /sdlc:incident	incident intent, thread summary	on-call	channel history	same

The two rules every command follows

1. A stage ends by committing an artifact; the next begins by reading it. Nothing progresses on a chat message alone. The commit is the handover and the audit record.
2. Artifacts never accept themselves. status: accepted is written by a person (or by the merge they perform). Findings never approve or block; hooks and branch protection do.

Tiered autonomy

Where	Default	How it is enforced
Interactive build	plan first, then edits; auto mode as guardrails mature	plan mode; hooks; CLAUDE.md verification block
Development environment	agent deploys freely	environments.development.autonomy: auto
Staging	agent deploys after checks, confirms	hook emits ask; CI manual job
Production	agent prepares; release manager authorises	production-gate hook (RELEASE_APPROVAL), CI environment approval, managed settings
Headless (CI) chains	continue or escalate decided by an independent agent	adversarial-reviewer confidence gate; needs-human label

This site is open source. Improve this page.