Courses › AI-Assisted Coding › The New Round

A Winning Workflow

Plan, delegate, verify, integrate — a loop that shows judgment, not just speed.

A reliable workflow keeps you in control of an AI-assisted session. The loop below works for building features, starting from scratch, or debugging — and it makes your judgment visible to the interviewer.

[Understand] -> [Plan] -> [Delegate a small step] -> [Verify] -> [Integrate] -> [Delegate a small step]

1. Understand (first 10–15% of the time)

  • Read the task carefully and list requirements and ambiguities. Ask the interviewer about the ambiguities — that's usually a deliberate part of the test.
  • For an existing codebase: run the tests first so you know the baseline. Skim the structure yourself; it's fine to ask the assistant to summarize the architecture, but verify its summary by opening the key files.
  • Identify conventions: how are handlers structured, how are errors returned, how are tests written? Your additions should match.

2. Plan

Say your plan out loud and write it down as a short checklist:

1. Add Task model fields: priority, due_date (validate formats)
2. Storage: persist new fields; migrate existing data
3. CLI: `add --priority --due`, `list --sort-by --filter`
4. Tests for parsing, sorting, filtering, invalid input
5. Refactor + README update

You own the plan, the data model and the module boundaries. This is where you show engineering judgment.

3. Delegate small, well-specified steps

Hand the assistant one step at a time, with context and constraints (see Prompting Under Pressure):

  • Small enough to review in a minute or two
  • Specific about files, function signatures and conventions to follow
  • Including the tests or examples that define "done"

Do some parts yourself — particularly the trickiest logic or the core data model — so the interviewer sees you code.

4. Verify — every time

  • Read the diff. Every line. Ask yourself whether you'd approve it in code review.
  • Run the tests after every step, not at the end.
  • Test edge cases the assistant may have missed: empty input, invalid formats, boundaries, concurrency.
  • When something fails, form a hypothesis first before asking the assistant to fix it. Blind "fix this error" loops waste time and hide your reasoning.

See Reviewing AI Code for what to look for.

5. Integrate and refactor

Generated code drifts in style. Keep the codebase coherent: consistent naming, no duplicated helpers, dead code removed, comments that explain *why*. Leave 10 minutes at the end for this and for a walkthrough.

Narrate throughout

Say what you're delegating and why, what you're checking, and what you found. "I'll have it write the argument-parsing boilerplate since it's mechanical — but I'll write the sorting comparator myself because tie-breaking on due dates is subtle." That one sentence shows the judgment the round is designed to measure.

[!TIP] Time budget for a 60-minute round Understand + plan: 10 min · Build in small steps: 35 min · Final testing, refactor and walkthrough: 15 min.