How Specification‑Driven Development Tames AI‑Generated Code with Three Loops and Six Human Checks

Specification‑Driven Development: Fixing What to Write

In a harsh‑environment industrial computer workstation (as of 2026‑07) the development process consists of six closed‑loop stages: requirement authoring, specification clarification, code generation, cross‑compilation, download verification, and issue fixing. The 需求.md file is the single source of truth referenced by all subsequent steps.

Clarifying requirements is the most error‑prone step. For a requirement such as “protect when hydraulic oil temperature is too high”, the AI must ask: what temperature threshold defines “too high”? What protection action (PWM reduction, alarm, shutdown) is required? Which task cycles are affected? Without these clarifications the model would make implicit assumptions.

Three Iterative Loops: Traceable Rather Than Waterfall

The workflow can be divided into three loops:

Requirement‑Specification Loop : Edit the requirement, the AI identifies vague points, asks clarification questions, and writes the answers back into the requirement, addressing the “cannot articulate” problem.

Generation‑Compilation Loop : The AI first produces specification documents, design notes, and task lists, then writes code. If the build fails, the system reads the build summary, feeds the errors back to the AI for automatic repair, repeating until a downloadable output/hec_app artifact is produced.

Verification‑Repair Loop : Before downloading to the device, on‑site personnel confirm safety conditions. After verification, any discovered issue is recorded and the cycle returns to repair or feature‑addition, ensuring feedback reaches the requirement and specification rather than becoming orphaned code.

Architecturally the key is not “AI is smarter” but that AI operates within a controlled chain: the requirement document and generated artifacts are centrally managed, and code generation is followed by build, artifact, and process‑state traceability. AI uncertainty is mitigated by documentation, specifications, build results, and human decision points.

Six Decision Points: AI Generates, Humans Guard

Six decision points require human confirmation:

Project/model selection

Requirement granularity

AI‑generated clarification assumptions

Scope of the current generation

Safety confirmation before download verification (the system prompts for a safety‑risk acknowledgment and requires the operator to type yes after confirming protective measures are in place)

Post‑verification judgment on whether to fix, add functionality, or deliver

Download verification is especially critical in industrial settings; the local documentation warns that the device will not run unless the operator explicitly acknowledges safety risks.

The threefold difficulty described in D01 cannot be solved by simply asking AI to write more code. The method requires knowledgeable on‑site personnel to articulate clear requirements, AI to operate consistently around a single source of truth, and every change to leave an auditable trace.

Specification‑driven development does not let documents dominate code; instead it enables code to answer the question “Why was this written this way?”