Interview scorecard template for technical interviews

Anand Chapla ·

An interview scorecard needs four things: the one question the round is there to answer, a short list of criteria, a written description of high, mid and low for every criterion, and a place to write down what the candidate actually did. The template below has all four, for a technical round where the candidate may use an AI assistant. Copy it and change the criteria to fit your role.

Most scorecard templates stop at the first two. They list criteria and a 1–5 scale, and leave each interviewer to decide privately what a 3 means. The two parts they leave out are the parts that make two interviewers’ scores comparable.

The template

Round: _____ · Role and level: _____ · Interviewer: _____ · Date: _____

This round is here to find out whether the candidate can: _____ (one sentence, written before the interview, e.g. “find and fix a bug in code they did not write”)

Criterion High (3) Mid (2) Low (1)
Understanding the problem Restates the task in their own words, asks about the unclear part, names an edge case before writing code Starts on the right problem, finds the edge cases later, while testing Starts building before reading the task; solves a different problem
Working with the assistant Gives it the constraint and the relevant code; reads what comes back; rejects or changes it when it is wrong Uses it for useful pieces, but accepts some output without reading it Pastes the whole task in and submits what comes back
Correctness and testing Runs the code early; writes or asks for a test that would have caught the bug; checks the empty and the odd case Runs the happy path; fixes failures when they appear Does not run the code, or calls it done while a test fails
Changes they made Every change to generated code has a reason they can name; removes what is not needed Keeps generated code mostly as it came, with some cleanup Cannot say what the generated code does
Explaining decisions Explains a choice during the work, and the explanation matches what they did Explains when asked, mostly matching the record Explanation does not match what happened in the session

For every criterion, write:

  • Score: 1, 2 or 3
  • Evidence: the moment that shows it — a time, a file and line, a message to the assistant, a test run. One line is enough.

Overall: strong yes · yes · no · strong no

One line for the debrief: _____

How to fill it in

Write the band descriptions before the first interview. If you cannot say what a mid answer to “working with the assistant” looks like, the criterion is not ready, and every interviewer will fill the gap with their own picture. The bands above are a starting point. Rewrite them in words that fit your codebase.

Write the evidence during the session, not after. A note like “14:22 — reverted the assistant’s fix after reading it, then added a failing test” takes ten seconds at the time and cannot be rebuilt from memory two hours later. A score with no evidence line next to it should count for nothing in the debrief.

Score before you talk to the other interviewers. Once one person says “strong yes” in the room, the others drift toward it. Each interviewer submits the card first, then the discussion starts.

Test the bands on real sessions. After your first three or four candidates, have two people score the same session separately. Wherever they are two points apart, the band description is unclear. Rewrite that row, not the people.

A filled-in row

Working with the assistant — 2 (mid). Asked the assistant to fix the failing date test with the error message and the function, which was good context (00:12). Accepted its second fix without reading it; the fix passed the test but changed the public function’s return type (00:31, src/schedule.ts lines 40–52). Did not notice until the reviewer asked.

(Illustrative — not a real candidate.)

The row says what happened and where. Anyone in the debrief can disagree with the 2, and they can disagree about the same moment.

Mistakes that make a scorecard useless

Only the top band is written. “Excellent problem solving” with nothing under it tells the interviewer what to look for and nothing about how to score anything less. The low band matters most, because that is the score you will have to explain.

Too many criteria. Past five or six, nobody watches all of them in one hour, so the extra rows get filled in from a general impression. Cut a criterion before you add one.

A “culture fit” row. It has no band anyone can write down, so it becomes the place where the interviewer’s general feeling goes. If you mean “explains their reasoning to a teammate”, write that.

Averaging away a low score. A 1 on correctness and 3s everywhere else averages to a pass. Decide in advance which criteria can sink a candidate by themselves, and write that on the card.

Scoring the final code only. When the candidate can use an assistant, working code is the expected result, not the evidence. The rows above score how they got there, which is why each one needs a moment from the session behind it.


Before your next round, write the low band for each row on your current scorecard. The row you cannot write is the one your interviewers are scoring by feel.

codesolara builds this card from the session itself: every score cites the moment behind it, and a second pass checks each citation. For what to grade when the candidate uses AI, see how to assess an engineer who uses AI.

// more