Drawing
THE MERGE WORKS
Fig. 1 — a machine whose working is the concept
A branch is a car on its own lane; main is the trunk line at the front. The loop runs from intent to merge and back: the Plan hopper releases a product-owner-accepted, tier-sized intent.md and feeds the queue, and the Maintain monitor writes each 2σ or 3σ breach back as a fresh intent. Every gauge is read straight off the machine. All six Anthropic AI-Native SDLC stages, simplified — the parts list marks what is the playbook's and what is local.
intent.md → queueintent.md → spec.md; weight by tier (local)plan.md → diff + tests| Part | Is really | Stage / term |
|---|---|---|
| Plan hopper | releases intent.md + spec.md at the rate a product owner accepts them (the accept is the merge of intent.md), feeds the queue | ①② Plan·Design · human accept |
| Intent card | planning weight — card height = tier | ADR · PR block · issue body (local) |
| A lane | one branch, one PR, one issue. Starts from a committed plan.md (plan mode); CLAUDE.md, skills and hooks shape the build but aren't drawn | ③ Build · worktree per session |
| Build lamp + eval | CI red→green. Changes to CLAUDE.md, skills or hooks also run the eval suite; each shipped breach fix adds an eval for the incident | ④ Test · feedback loop + continuous evals |
| “AI review” tag | agentic review of every PR against REVIEW.md; findings inform, never approve | ⑤ Deploy · layered agentic review |
| Reviewer beam | human code-owner review + merge approval | ⑤ Deploy · human for high/med; low self-merges (local deviation) |
| The fitter | rebase onto the current trunk tip | local extension — the playbook doesn't cover rebase |
| Crane | the merge — locks car onto trunk | ⑤ main advances |
| Maintain monitor | deterministic control bands: 1σ log, 2σ read-only diagnosis, 3σ may propose a PR or pre-approved runbook; each 2σ/3σ diagnosis is written as intent.md | ⑥ Maintain · post-merge signals, not prod |
A merge is the lock-on. Until the crane lifts it, a car is a reversible proposal on its own lane. The crane is the one irreversible step — the commit becomes shared history everyone branches from. And the trunk moves while you work: every merge forces the other in-flight cars through the fitter (rebase) before their own lock-on. That reconciliation is the tax parallel work pays.
Now the loop closes. The Plan hopper keeps the queue topped to lanes×2 (queue health), stamping each intent at its tier weight — a high-risk car carries a tall ADR stack, a low-risk one a thin slip. Every merge pulls the next intent. And the Maintain monitor watches post-merge signals — nightly CI, an eval regression, cost drift, a trace — against deterministic control bands. 1σ is logged; at 2σ Claude diagnoses read-only, and at 3σ it may also propose a PR or a pre-approved runbook. Either way the diagnosis is written as a fresh intent.md that jumps to the front of the queue (a local priority rule). That is the playbook's “a breached control band in production writes the next intent.md.” See the control-band figure for how detection, tiers and triage work.
Honest adaptation: the playbook gates at merge (code-owner approval) and at release (an approval hook, with dev/staging/prod autonomy tiers). This machine stops at merge: there's no release gate because there's no production deploy yet, so the return-pipe draws from post-merge signals instead of a prod environment.
The ceiling isn't the machine — it's the beam. Pull Lanes to 4–5. Builds still go green (open→green holds), but green→merge climbs and cars start to HOLD: one reviewer covers WIP 2, and faster models just feed more cars into that queue. The beam is fixed; the share of cars that need it is what you tune. That is the playbook's move — layered agentic review on every PR, “human review reserved for regulated and critical code” — so at 5 lanes, flip Needs human from 85% to 30%: green→merge drops and merges per minute rise, with the same single reviewer. Push it too low and you're shipping under-reviewed, which is the other half of the same warning. The target is verified merges per week: add lanes only while the beam still reaches every car that needs it.
DRAWN IN THE SPIRIT OF THE “ANATOMY” SKILL — one model drives both the geometry and the readouts · isometric · light + dark · interactive. Traces all six stages of Anthropic's AI-Native SDLC Playbook.