Note-taking and post-interview summaries · AI for HR & Recruiters

You finish a sixty-minute interview, walk the candidate to the door, sit down at your desk, and the phone rings. By the time the call ends twenty minutes later, the candidate’s answer to the second scenario question — the one that actually told you something — is gone. You will remember a feeling. You will not remember the words. The decision meeting tomorrow will be a comparison of feelings, and feelings are exactly what the structured rubric you spent yesterday building is supposed to protect you against. This section is about the one workflow that closes that gap: a fast, structured AI summary written from your own rough notes, against the rubric, before the conversation cools.

The 90-second window

The most valuable ninety seconds in the entire hiring process are the ones immediately after the candidate leaves the room. The interview is still in your short-term memory in vivid detail. Specific phrases the candidate used are still recoverable. Your impression of how they handled the second probe on the conflict question is still distinct from your impression of how they handled the first. Five minutes later, all of this begins to blur. By the next morning, what survives is a sentiment — “I liked her,” “He felt off” — and a few headline moments. The rubric-shaped detail is gone.

The discipline is straightforward and is mostly about not letting yourself answer the phone yet. Before you do anything else, type rough notes against each competency on the rubric for two or three minutes. Not polished prose. Bullet points. Direct quotes you remember. The moments where you wrote something on the printed rubric during the interview, expanded into a sentence. This is the raw material the model will work with. If it is not captured here, no amount of clever prompting later will recover it.

Raw notes in, structured summary out

Once the rough notes exist, the AI workflow has one job: turn them into a structured summary against the rubric, with evidence quoted from the notes themselves, and nothing invented. The prompt pattern that does this reliably looks like this:

“Here are my rough notes from a 60-minute interview with a mid-level backend engineer candidate. The four competencies on the rubric are: backend system design, code quality and debugging discipline, ownership of incidents in production, and clear written communication with non-engineers. Produce a structured one-page summary with four sections, one per competency. For each: a strengths paragraph, a gaps paragraph, and at least one direct quote from my notes as evidence. End with a recommended next step (advance, second interview, decline) and a one-sentence justification. Use only what is in the notes; do not invent evidence. If a competency was not tested, say so explicitly.”

Five things make this prompt work. One, the rubric is named in the prompt, so the model knows the headings to organise against. Two, the structure is fixed — strengths, gaps, evidence — so every candidate’s summary has the same shape and decision meetings can compare them directly. Three, the “evidence quoted from notes” instruction forces the model to ground every claim in something you actually wrote, not in its general sense of what a good backend engineer looks like. Four, the “do not invent evidence” instruction is explicit. Five, the “if a competency was not tested, say so” clause surfaces gaps in the interview itself — which is information the panel needs before deciding.

The evidence-only instruction is doing real work

The single most important sentence in that prompt is “use only what is in the notes; do not invent evidence.” Without it, the model will fill in plausible-looking detail — invented examples, invented strengths, invented gaps — because that is what its training data tells it a good interview summary looks like. The summary will read fluently and will quietly contain claims about the candidate that did not happen in the room.

This is the same hallucination pattern from every other section of this course, but it is more dangerous here because the output is going to influence whether a person gets a job. A confidently written invented strength can swing a borderline candidate into the shortlist. A confidently written invented gap can knock a strong candidate out. The evidence-only instruction does not eliminate this risk; it sharply reduces it. The second safeguard is your read-through: before the summary leaves your desk for the decision meeting, read every quoted piece of evidence and confirm it is in your notes. If a quote is in the summary but not in the notes, the model invented it, and the rest of the summary now needs re-reading with that suspicion.

Bias prevention: the comparability that the same format buys you

The decision meeting after a round of interviews is the moment where the structured-interview discipline either pays off or collapses. If every candidate’s summary follows the same four-section, strengths-and-gaps-with-evidence format, the panel is forced into a comparison the rubric was built for. If half the candidates have a structured summary and half have a “great chat, she’s our person” paragraph, the conversation drifts to whichever candidate the loudest panelist remembers most warmly — and the structure was wasted.

The same format across every candidate is itself the bias-reduction mechanism. Without it, decision meetings reliably drift toward likeability, recency, and resemblance to the panel. With it, the panel still has to make judgement calls, but the calls are against comparable evidence. A candidate who scored a 4 on ownership has an evidence paragraph saying why; another candidate who scored a 3 on ownership has an evidence paragraph saying why; the panel can disagree with either score, but they cannot pretend the comparison was not made.

Confidentiality and which tool you use

A practical detail that matters in Nepal: most HR teams are using public chatbot tools (ChatGPT, Claude, Gemini) rather than enterprise deployments. Before pasting interview notes into any public tool, redact the candidate’s name, contact details, current employer, and any other identifier. Replace with “Candidate A” or initials. The notes are confidential professional information; the candidate did not consent to having their words processed by a model whose retention policy you cannot audit.

This is not a hypothetical concern. A candidate’s name plus a “gaps” paragraph entered into a public tool, then later surfaced in some indirect way, is the kind of incident that ends an HR career. The redaction takes thirty seconds. If your organisation has an enterprise deployment with a no-training data-handling agreement, this concern is reduced but not eliminated — even there, professional discretion says you keep identifiers out of prompts unless they are necessary, and for a summary against a rubric they are not necessary.

A realistic rhythm for a busy day

A team that runs four interviews in a day cannot spend an hour on each summary, and does not need to. A workable rhythm: ten minutes of rough notes immediately after the interview, two minutes to run the structured-summary prompt, three minutes to read the summary against your notes and fix anything that drifted, send to the panel folder. Fifteen minutes per candidate, sixty minutes for the day. Without this rhythm, the same hour is spent in the decision meeting the next morning trying to reconstruct what happened, badly.

The hardest part is not the prompt or the tool. It is the habit of not letting yourself check the phone for ten minutes after the candidate leaves. The recruiters who consistently make better hires are usually the ones who guard those ten minutes the most aggressively.

Check your understanding

Quick check

—

A recruiter has just finished a 60-minute interview and has scattered handwritten notes on the printed rubric. The decision meeting is tomorrow morning. What is the best use of AI in the next fifteen minutes?

Ask the model to imagine what the interview probably contained based on the JD and produce a summary from that.
Skip the summary — the panel will discuss the candidate tomorrow anyway.
Type up the rough notes (with the candidate’s name redacted), then prompt the model for a structured summary against the rubric with strengths, gaps, and direct quotes from the notes as evidence, ending with a recommended next step.
Paste the candidate’s CV into the model and ask it to predict how the interview went.

What comes next

You have the kit, the rubric, the calibrated panel, and the structured summaries. The candidate is moving toward an offer. The next section is the step almost every team is tempted to skip when the AI score is high and the interview was strong — the reference check — and why that temptation is exactly when references catch the most.