# AI coding tools: working checklist

Companion to https://lunaprompts.com/blog/ai-assisted-coding-assessment-guide

LunaPrompts Editorial. Original illustrative exercise.

## Define the workflow

- Role or campaign:
- Owner:
- Decision this workflow should support:
- Candidate instructions and support route:

## Worked example: Review an AI-generated API change

Give the candidate a small endpoint with a missing validation rule. Supply the expected error behavior and an existing success response. State which assistant is available and make sure access works before the timer begins.

The candidate can use the assistant to propose a patch. During review, look for evidence that they inspected the project context, constrained the change, and read the resulting diff. A generated patch that fixes validation but renames a response field introduces a useful discussion about interface ownership.

Ask, “Which behavior did you preserve, and how did you check it?” A strong explanation can point to a request, an assertion, or a focused diff. If the answer is simply that the assistant ran tests, ask which tests and what they establish. Existing tests may not cover the changed contract.

Keep the assessment comparable. Give candidates the same relevant repository state, instructions, time allowance, and tool access. Do not silently reward access to a better model or an already configured personal environment. When the role requires a specific tool, explain that requirement; when it does not, keep the rubric focused on engineering decisions.

## Review checklist

- [ ] Context: Inspects the relevant files and behavior
  Evidence or open question:
- [ ] Direction: Gives the assistant a bounded task
  Evidence or open question:
- [ ] Review: Finds unintended changes in the diff
  Evidence or open question:
- [ ] Verification: Checks both the new requirement and preserved behavior
  Evidence or open question:

## Working session

Choose a small bug in an example repository. Write one acceptance criterion and one behavior that must remain unchanged, then use both in the review.

## Avoid this mistake

A polished prompt is not evidence that the resulting code is correct. Review the work it produced and the candidate’s verification.

## Record the next step

- Observed evidence:
- What remains uncertain:
- Person responsible:
- Next action and date:

Keep candidate examples anonymous when sharing this worksheet.