ai‑sdlc · simulation

Software that ships at agent speed, with people at every gate.

This is the lifecycle your engineering team is adopting, running live. Claude writes the files (an intent, a spec, a plan, the code, a review, an incident report). A named person accepts each one at an amber gate. When something is wrong, the stage that owns it fixes it before anything moves forward, and the same mistake is not made twice. Drag to look around; click a station to zoom in.

Your browser could not start 3D graphics, so the simulation cannot render here.

The written version below covers the same loop.

Stage 1 · PlanClaude works

An idea arrives

Someone in claims operations describes a problem in their own words.

Who decides: nobody yet
intent.mdsdlc/changes/…

      
Drag to orbit · pinch or wheel to zoom · A card moving backwards is work sent back to the stage that owns the fix. Keys: space pause · F full screen · 1–6 focus · 0 overview

What you are watching

Five kinds of thing appear on the ring. Everything else is scenery.

A paper card

An artifact Claude wrote: intent.md, spec.md, plan.md, the pull request, the incident. It is a real file saved in the repository, in your words.

An amber arch

A gate. The card stops until a named person accepts it: product owner, engineer, code owner, release manager, service owner. Claude cannot pass a gate by itself.

A teal shield

A hook: a small script that runs automatically while Claude works. Protected files, secrets, tests during a bug fix, and production releases are guarded this way.

A red block

A mistake caught in the act, such as editing generated code or committing a secret. It hits the shield and bounces, and it is fixed in that same stage before anything moves on. The second time, the correction goes into CLAUDE.md so it stops happening.

The growing chain

Every accepted artifact and every approval becomes a commit — teal for one Claude wrote, amber for one a human made by approving or authorising, the same colors as the arch and the shield above. The chain in the centre is the audit trail. Click or tap any dot to see what it was.

The six stages, and who decides

Each stage ends by saving one file. The next stage starts by reading it. Nothing moves on a chat message alone.

StageClaude producesA person decidesCommand
1Planintent.md — the problem, the desired outcome, constraints, open questions, in the originator's wordsProduct owner accepts by merging
2Designspec.md — requirements and design under the organisation's security, compliance, brand and UX policies, with flagged concerns firstProduct owner accepts; policy owners decide each flagged concern
3Buildplan.md then the code, with tests run after every step; the plan is kept honestEngineer accepts the plan before any code; tech lead if high risk
4TestProof: build, tests, lint, screenshots, attached to the pull request; a fresh-eyes verifier re-checksAutomatic, tests cannot be weakened during a fix
5DeployAn AI review of every pull request (bugs, security, compliance against the spec and plan), then the releaseCode owner approves; release manager authorises production
6MaintainA script watches production; on a real anomaly Claude diagnoses and writes a new intent.mdService owner triages: fix now, schedule, dismiss

Always available (not stage-scoped):

The teams at each stage

Nobody works alone at a stage, and Claude is a member of every team, not a replacement for one. The names below are the ones you see acting in the simulation. As someone acts, their marker on the platform lights up.

What changes for product

You review, you don't transcribe.

An idea becomes a committed intent in hours, in the originator's own words. The spec arrives written; your job is to read the flagged concerns first, settle them with their owners, and accept. No requirements workshop.

Your decisions are the record.

Accepting is a one-click merge with your name and the time on it. Nobody has to write minutes; the repository is the audit trail, and /sdlc:status answers "where is it?" without a meeting.

The rules apply while the work happens.

Security, compliance, brand and UX policies shape the spec as it is written and are enforced by hooks as code is built. Conflicts are flagged for you instead of discovered in a review weeks later.