A safer AI-assisted delivery loop

AI is most useful in delivery when it shortens the distance between a clear problem and reviewable evidence. It is least useful when it becomes an authority: deciding scope, changing multiple concerns at once, or declaring its own output safe.

The distinction is not philosophical. It changes the shape of the workflow.

Start with a bounded brief

A useful brief names the behaviour to change, the boundary not to cross and the evidence required before review. In the accompanying silent demonstration, the fixture contains three explicit checks: contract, test and policy.

Those checks exist before the patch. The model does not invent the acceptance criteria after seeing its own implementation.

goal       reject invalid input before persistence
boundary   synthetic fixture data only
evidence   contract + unit test + policy check

This is deliberately smaller than “improve the service”. A narrow brief makes failure legible.

Separate discovery from change

The first pass should explain the repository as it exists: where behaviour enters, where it is tested and which constraints already apply. The output is a plan another engineer can challenge before code changes.

Then make the smallest coherent patch. Implementation and its behavioural test travel together. Unrelated cleanup stays out, even when it is tempting.

Verification is a pipeline, not a feeling

The demonstration reports three deterministic results:

1. the unit suite passes;
2. linting passes;
3. the policy check passes.

In a real repository those commands are project-specific, and a passing result is not proof of correctness. It is evidence attached to a review. The reviewer still decides whether the brief was right, the tests are meaningful and the change belongs in production.

Produce a review packet

The final artifact should be easy to inspect without replaying the whole session. A compact review packet records:

• the requested behaviour;
• files changed and why;
• commands run and their results;
• assumptions or unresolved risks;
• the exact diff awaiting approval.

That packet is the handover boundary. AI may have accelerated discovery and drafting, but the evidence is conventional engineering evidence.

The operating rule

Use AI to compress toil, not accountability. Define the checks first, keep the patch bounded, and make the final decision from artifacts that do not require trusting the model that produced them.

The source-controlled VHS tape behind the demonstration uses only synthetic fixture paths and deterministic output. It exists to show the sequence, not to pretend a twenty-five-second recording is a production benchmark.