You have a design review tomorrow and little time to experiment. Asking an assistant to 'check my chip' gives it an undefined problem and gives you no clear way to judge success. A better first use is a bounded task whose answer you already have the means to inspect: explain a short module, organize a synthetic log, or draft review questions. The assistant helps produce an artifact; the engineer decides whether the artifact is correct and useful.
A prompt is simply the request and supporting material you give the assistant. A response is generated text, code, or another draft. A fluent response can still include a false statement. Choose a task with a small input, an explicit output, and an inexpensive check. Define success before sending the prompt. For an explanation, success might mean every behavioral statement points to a branch or signal in the supplied source.
Discipline | Suitable first task | Unsuitable conclusion from a response alone |
|---|---|---|
RTL | Explain a synthetic counter and list its boundary cases. | The block is correct for every use or meets frequency. |
DV | Draft reset and enable scenarios for review. | The generated scenarios prove coverage closure. |
STA | Explain fields in a short, invented timing report. | A real path is safe to waive without constraints and analysis. |
PD | Turn a synthetic congestion note into questions to investigate. | A placement change will fix routing or improve silicon results. |
Synthetic request, no model call performed:
Given a four-bit counter that resets to zero and increments when enabled,
draft three review questions. Use only these facts. Do not claim sign-off.
Synthetic response for assessment:
1. Does reset take priority when enable is also high?
2. Does the counter hold when enable is low?
3. Does incrementing fifteen wrap to zero?These are useful questions because each leads to a concrete check. The description does not say whether reset is synchronous or asynchronous, so the answer should ask for that detail instead of inventing it. Notice the difference between proposing 'check wraparound' and declaring 'wraparound passes.' The first is a plan; the second requires evidence that has not been supplied. Even a good question list may miss cases, so compare it with your own reading before using it in a review.
Set a stopping point as well. Spend ten minutes producing a short question list, then inspect it yourself. If the assistant repeatedly needs information you cannot share, replace the task with a synthetic example or work locally under approved policy. Repeatedly rephrasing a request cannot supply a missing specification. A useful result may be a precise question for the block owner, rather than an answer.
The explanation and checklist are suitable drafting tasks with review. The CDC certification is unsupported: a name gives no clock crossings or synchronization evidence. Rewrite it as 'List the clock, reset, crossing, and CDC-report information a reviewer would need before assessing this block.'
For your own task, write one sentence each for input, output, check, and stopping condition. An acceptable DV example supplies a tiny interface, requests five candidate scenarios, checks them against that interface, and stops after a reviewed list. It does not ask for a coverage percentage without a coverage model and results.
Drafting assistance cannot replace design intent, authorized data access, or sign-off evidence. NIST AI 600-1, Generative AI Profile (July 2024), section 2.2, describes confidently incorrect generated content as confabulation. In this course, attaching evidence to each claim is the practical response to that risk.