Here's the situation I was trying to fix.
I bombed a final-round interview. Not catastrophically—no silence, no disaster—but I walked out knowing I'd given two weak answers and couldn't quite say why they were weak. By the time I got home, the details had already blurred. I remembered the feeling of doing badly and almost none of the substance.
So I couldn't fix anything. I could only resolve to "do better next time," which is not a plan.
This is the workflow I built after that. It's short, it runs on the same day as the interview, and its job is to convert a fuzzy sense of "that didn't land" into a specific, repairable list. It won't make you a better interviewee by itself. It will tell you exactly what to practice.
Why Post-Interview Reflection Fails
Three reasons, and none of them are about effort.
Memory Decays Fast and Selectively
Within a few hours, the content of an interview fades and the emotion stays. You remember you felt bad about the systems-design question, but not what you actually said. The felt sense is vivid; the transcript is gone. This is why "I'll think about it on the drive home" produces nothing usable.
You Can't Audit Yourself From Inside
Two problems make self-assessment unreliable in the moment:
You don't know the standard. You're comparing your answer to an imagined ideal, not to what the interviewer was actually listening for.
You're emotionally compromised. The question that felt like a disaster may have been fine; the one that felt smooth may have missed the point entirely. Intensity is a poor signal.
Most Advice Is Generic
"Use the STAR method." "Prepare stories." "Ask good questions." All true, all useless after the fact, because none of it tells you which of your answers was weak or what specifically to change.
The fix is to produce a specific artifact—not a feeling, not a resolution, but a list of named weak answers with a diagnosis for each.
The Workflow
The whole thing runs in about forty minutes on the day of the interview. It has four steps, and one of them is where AI does real work.
Step 1: Dump the Interview Immediately
Within an hour of finishing, while the details are still present, write down everything you can remember. Don't edit, don't judge, don't organize. Just get it out.
Capture:
Each question as you remember it, in the interviewer's words if possible
What you actually said, roughly, in your own words—not what you wish you'd said
Where you felt uncertain, rambled, or guessed
Where you felt strong
Any follow-up questions the interviewer asked, since those often reveal what they were probing for
The critical discipline: write what you said, not what you should have said. The moment you start improving the answer, you've lost the data you need. This is a record of the interview, not a rehearsal.
Length is fine. More raw material is better at this stage.
Step 2: Identify the Weak Answers
Not the ones that felt bad. The ones that were structurally weak. A quick way to spot them in your dump:
No outcome. You described a process but never said what changed or what result followed.
No specifics. You spoke in generalities—"we improved the process"—without names, numbers, or a concrete situation.
No ownership. You said "we" throughout and never made your own contribution clear.
No point. You answered a question but never connected it back to what the interviewer was probably asking.
No end. You rambled, or stopped without landing the answer.
Flag every answer that has one or more of these. Don't try to fix them yet—just name them.
Step 3: Run the AI Diagnosis
Now AI earns its keep, and this is the step most people get backwards. Don't ask the model to rewrite your answers. Ask it to diagnose them.
A prompt that works:
Below are my raw notes from a job interview, including the questions I was asked and roughly what I said. For each answer, identify the specific weakness: missing outcome, missing specifics, unclear ownership, rambling, no point, or something else. Do not rewrite my answers. Do not tell me what I should have said. Only diagnose the weaknesses and explain what the interviewer was likely listening for.
Then paste your Step 1 dump.
A useful model response names the problem for each answer and explains the gap between what you said and what the question was probably testing. This is the key move: diagnosis, not repair. Rewriting comes later, and only for the answers that matter.
One caution: the model doesn't know the room. It can spot structural weakness—missing outcome, no ownership—but it can't know that a particular company cares more about one dimension than another. Treat its diagnosis as a hypothesis, not a verdict.
Step 4: Build the Repair List
From the diagnosis, produce a short list. For each weak answer:
The question
The weakness (from Step 2 or 3)
The missing piece (the number, the outcome, the specific situation you should have named)
The replacement story or data point you already have, or need to find
This list is the entire point of the workflow. It's what you practice before the next interview—not the interview in general, but these specific answers, with these specific gaps filled.
Cap the list at three to five items. More than that and you won't practice any of them properly. If the diagnosis turned up ten weak answers, the top three are what matter.
The Before and After
Here's a real pair from that first debrief.
Before (raw note):
They asked about a time I handled competing priorities. I talked about the migration project, how I worked with the team to sequence things, and we managed to get it done. Felt okay but I noticed they asked a follow-up about what I personally decided.
After (diagnosis):
Question: A time you handled competing priorities.
Weakness: No ownership, no outcome. Answer stayed at "we" throughout, and the follow-up question shows the interviewer was hunting for your individual decision—which you never made explicit.
Missing piece: A specific moment where you chose what to deprioritize, and what happened as a result.
Repair: The week you cut the reporting feature to protect the migration deadline. You made that call. Name the tradeoff and the consequence.
The before note is a feeling. The after note is a repair. And the second one took about three minutes of honest diagnosis.
What Still Needs You
The parts of this workflow that don't get automated:
Writing what you said, not what you should have said. Honesty here is the whole game. If the dump is aspirational, the diagnosis is worthless.
Judging which weaknesses matter. The model will flag everything. You decide which three answers are worth repairing before the next interview, based on the role and the company.
Knowing what you actually have. The model can suggest a missing piece—a story, a number—but only you know whether it's true and whether you can tell it convincingly.
Sitting with the discomfort. Reading your own weak answers is unpleasant. The workflow only works if you're willing to do it.
The model diagnoses. You decide what to fix and then go fix it.
The Forty-Minute Version
If you have an interview coming up—or just finished one:
Within an hour, dump everything you remember. What you said, not what you should have said.
Flag answers with no outcome, no specifics, no ownership, no point, or no end.
Run the AI diagnosis prompt. Ask for weaknesses, not rewrites.
Build a repair list of three to five answers with the missing piece for each.
Practice those specific answers before the next interview.
One debrief won't transform your interviewing. But a debrief after every interview compounds fast—because each one tells you exactly what to fix, instead of leaving you with a feeling and a resolution to "do better."
Better work first. More options next.
Make the workflow earn its place.
No feedback yet — submit the first.