AI-assisted quantum programming needs a draft, verify, approve loop.

QFlow describes an AI Builder that can draft circuits, route notes, source changes, and evidence summaries from a user request. This can reduce the blank-page cost of starting a Bell, Grover, QAOA, or custom workflow and help a mixed team translate an objective into inspectable artifacts.

Every artifact remains a draft until reviewed. A plausible circuit can encode the wrong oracle, reverse a control, omit a measurement, misuse a parameter, or target an unsupported operation. Fluent language and a clean diagram are not evidence of quantum correctness.

The request should include the objective, input size, expected behavior, simulator or provider preference, source format, and required evidence. It should also state what the assistant must not assume: provider entitlement, unrestricted qubits, arbitrary Python execution, or a claim of quantum advantage.

Ask for the smallest simulator-first draft and an explanation of each block. This makes defects easier to locate and establishes a reference before provider routing. Complex workflows should be decomposed into reviewable components instead of generated as one opaque circuit.

Review register sizes, initial state, gate order, controls and targets, parameters, measurements, and classical-bit mapping. Algorithm-specific checks follow: a Grover draft needs the intended oracle and diffusion path; a QAOA draft needs a declared cost model, mixer, parameter boundary, and classical objective.

Test the smallest meaningful instance against an analytic expectation or strong classical reference. A simulator result that disagrees should stop the workflow before hardware routing. Agreement is necessary evidence for the draft, not proof that an algorithm will scale or outperform a classical method.

QFlow keeps generated source visible beside the workflow. Review imports, registers, operations, parameters, measurements, and framework-specific behavior. Tie the exact executed source snapshot to the run instead of inferring it from the latest editor state.

Generated code deserves normal software checks: version review, deterministic unit cases where possible, dependency scrutiny, and comparison with official SDK behavior. QFlow’s published boundary keeps raw Python off by default, so an assistant should not silently widen a circuit task into arbitrary code execution.

An assistant may suggest a simulator or provider fit. The team must verify adapter status, operations, target access, credentials, region, quotas, queue, expected charges, and why hardware is needed. QFlow’s public matrix makes clear that routes have different implementation levels.

The approval checkpoint should display the workflow version, source snapshot, route, target, shots, and known resource or account implications. A failed preflight remains visible; it must not be silently converted into another execution mode. Provider terms and access remain independent from QFlow.

Check an AI-written summary against counts, trace, source, route, exports, and reviewer notes. The assistant must not turn finite-shot variation into a causal explanation, call simulation hardware evidence, omit failed attempts, or convert a logical algorithm claim into a runtime advantage.

People also approve shared results. The packet exposes proof required for review while excluding provider tokens, billing state, administrative notes, private users, and raw credentials. AI can help draft the narrative; the artifacts and manifest determine what the narrative may claim.

A useful evaluation set includes reversed controls, missing measurements, invalid parameters, unsupported operations, ambiguous provider access, misleading performance language, and summaries that omit negative runs. Measure whether the system surfaces uncertainty and whether reviewers correct the defect before submission.

QuanBench+ and Qiskit HumanEval examine functional correctness in generated quantum code, reinforcing a practical point: fluent code generation needs task-specific execution tests and human review. A trustworthy assistant makes the workflow more inspectable; it does not hide judgment behind polished output.