AI Tools Are Commodity. The Spec Layer Is Not.

Porting a spec-driven AI workflow from Claude Code to Copilot — and what stayed the same.

AI slop dropped sharply on my 30K-LOC codebase once I stopped treating Claude as the deliverable and started treating the spec artifacts as the deliverable.

I’m now porting the same harness to Copilot (with Claude as the model), and the migration is almost entirely a configuration exercise — not a re-architecture.

The flow

·      Refine: every issue passes through advisor consultation (SEO, UX, content-strategy) before planning

·      Consolidate: refinement outputs merge into a single LOCKED-SPEC artifact

·      Plan: implementation plan reads the spec; locks architectural decisions explicitly

·      Implement: agent writes against the plan, with file-level test mapping

·      Code review: code-quality gate

·      Audit: spec-compliance gate. Cross-checks every binding commitment against the PR diff. Drift surfaces with three resolution paths: fix code, amend spec, or signed override.

Code review catches code issues. Audit catches silent narrowing — the failure mode where an agent ships something that compiles, passes tests, and quietly dropped a commitment. The audit is where the slop reduction actually came from.

Porting to Copilot has been mostly mechanical — the spec artifacts don’t care which front-end produces them. Front-ends are commodity. The durable investment is the artifact layer and the gates that enforce it, not whichever tool you license this quarter.

But none of this replaces human judgment. It concentrates it.

Three pivots that stay irreplaceably human:

1.        The initial ask. What to build, what tradeoffs to accept, what feels right for the product. No spec discipline rescues a bad request.

2.        Reviewing the spec. Someone has to read the LOCKED-SPEC and decide it’s actually right — not just complete. The audit checks the implementation against the spec. It doesn’t check that the spec was good.

3.        Reviewing the PR. Audit clean, tests green — the judgment call about whether the change is the right shape, the right scope, the right time is still yours.

Everything between those three moments is mechanical. The human pivots are where platform discipline should concentrate attention.

Curious whether others are running similar disciplines

·      Is your AI workflow producing artifacts your platform owns, or just edits in a tool you rent?

·      Where does your spec-compliance audit live, separate from code review?

·      Has anyone done a Claude Code → Copilot port at scale? What broke, what carried over?

#PlatformEngineering #DeveloperProductivity #AIAssistedDevelopment #SpecDrivenDevelopment