AI-Native SDLC: From intent.md to bands.yaml Governance
Long-form guide from Anthropic by Louis Claxton (Applied AI team), published on August 21, 2026 on the claude.com blog: a stated 40-minute read, roughly 64,000 characters, presented as a collection of plays drawn from the team's work with its clients. (A) The diagnosis: with code no longer the bottleneck, it shifts to the stages on either side of the build (plan, review/test, deploy), line-by-line controls stop holding once the agent writes most of the diff, and governance cost rises as exceptions still route through periodic committees. (B) The response: six stages (Plan, Design, Build, Test, Deploy, Maintain) organized as a loop rather than a chain, each ending with a committed artifact that the next stage reads — intent.md, spec.md, plan.md, the diff and its tests, the PR and its findings, the incident record. (1) Institutional knowledge becomes versioned files: CLAUDE.md, skills, REVIEW.md, bands.yaml. (2) Governance splits into two layers, with the skill positioned as an advisory control and the hook as the deterministic layer behind it.
By Louis Claxton// Source claude.com ↗/Reading 2 min/.md// Auto-verified translation
Louis Claxton, of Anthropic's Applied AI team, published an implementation guide for an "AI-native" software development lifecycle on August 21, 2026. The starting point is an imbalance: organizations now write code at a speed unimaginable a year earlier, but the processes around it — approval gates, reviews, handoffs, policies — haven't moved. The traditional SDLC was designed for a world where writing code was the longest and costliest stage; its controls also assume that every action is taken by a human.
Three consequences follow. The bottleneck shifts to the stages that still run at human speed, on either side of the build. Controls stop being applicable: reading every line made sense when a person had written it. And governance cost rises, as exceptions route through periodic committees.
The response keeps the control objectives and changes how they're executed. The process becomes a loop, with AI embedded at every point, organized into six stages — Plan, Design, Build, Test, Deploy, Maintain — broken down into plays that all follow the same grid, down to the metrics. The throughline is the committed artifact. Intent is captured by its original author as intent.md; requirements and design merge into a single session producing spec.md, constrained by the brand, security, compliance and UX skills; the build starts in plan mode and locks plan.md before any code is written. The commit chain serves as the audit trail.
Institutional knowledge becomes versioned files: CLAUDE.md for repository context, skills for cross-cutting policies, REVIEW.md for review doctrine, bands.yaml for production thresholds. Governance splits into two layers, with the skill as an advisory control and the hook as the deterministic layer that blocks or requests approval. A managed settings example details, key by key, what each setting buys in terms of control, from refusing to read secrets to enforcing a minimum version floor.
The Maintain stage closes the loop: a deterministic script monitors a metric, and crossing a band invokes Claude with no human in the call path, at an autonomy level set by the tier. What the agent finds is rewritten as intent.md and fed back into the cycle. Claude Tag, in public beta on Slack, extends the pattern to incidents arriving via chat. No quantified results are put forward: the guide provides metrics to measure and names their source.
Key takeaways
Three consequences once the build stops being the constraint. (1) the bottleneck shifts to the stages that still run at human speed (plan, review/test, deploy); (2) controls become inapplicable — reading every line made sense when a human had written it; (3) governance cost rises, as exceptions route through periodic committees. Example given: a security team sized for human throughput, facing which either the review queue grows or code ships under-reviewed.
The throughline is the committed artifact. , not the tool: each stage ends by writing to version control, the next begins by reading it, and the commit chain is the audit trail. .md dominates upstream because the product owner and the agent read the same file; from the build stage onward, the artifact is the code and its traces.
Cascading triggers. an accepted intent.md triggers the requirements/design pass, an approved spec.md triggers plan mode, a merged PR triggers the pipeline, a band crossed in production writes the next intent.md. Teams start by prompting each stage by hand; the target state is the loop where each accepted artifact arms the next gate.
Skill vs. hook — the distinction carries the whole control edifice. the skill makes policy compliance likely without forcing a session to comply; the hook is deterministic and blocks the action. A policy that must always hold needs a hook or a review pass behind the skill. Corollary: a hook that requests human approval belongs in deployment, not in the build, where it would put a person back on the critical path of every parallel session.
Legacy systems. for each artifact, name one system as the source of truth (the repo, or Jira/ServiceNow with the .md files as working copies), with everything else holding only a link. Simple chaining — the artifact carries the record ID, the record carries the commit SHA — is offered as the minimum starting bar.
Test. the feedback loop (tests, build, screenshot diff) runs throughout the task; the verifier subagent is a final pass with fresh context, so the verdict isn't colored by the assumptions that produced the code. For a fix, write the failing test first, commit it, then block the agent from editing it via a hook. Evals are the AI-native counterpart to QA gates: 20 to 50 real tasks replayed on every change to CLAUDE.md, a skill, or a hook, with every incident becoming a permanent eval.
Maintain, closing the loop.detection stays deterministic (rolling-window mean and standard deviation, Western Electric rules, a versioned and tested script, no model involved); Claude is invoked only once a band is crossed, and the tier fixes what it can do — 1σ logs, 2σ diagnoses read-only, 3σ proposes (a PR or a pre-approved runbook). Rollback is named as the path that must be the most rehearsed in the pipeline.
⚠️ What the piece doesn't quantify: no quantified results, neither time savings nor adoption rate. The guide's numbers are implementation parameters (20-50 evals, two or three parallel sessions to start); outcomes remain metrics to measure yourself, with their source named each time (git log, PR metadata, OpenTelemetry export, DORA, incident-tracking tool).
To connect. [[sfeir-sdlc-ia-cycle-11-phases-2026-06-16]] (a competing breakdown, into eleven phases) and [[sfeir-code-review-anneau-contraintes-2026-07-30]] (the ring of constraints around the agent, where hooks and review are here two distinct rings).
Key figures
delivery performance, a lagging indicator of the CI/CD play