When you practice interviews with AI, ask for short feedback, not rewrites.
What the Reel says
When you practice interviews with AI, ask for short feedback, not rewrites.
I practiced for an outdoor shop job.
3 sentences
my first answer
It also rewrote my answer twice.
They added duties I can't back up.
So I said: one question, wait, two-line feedback.
It gave one fix, then the next question.
So you do the talking, not the AI.
Sources and assumptions
- Prompting best practices (Claude Platform docs), section 'Be clear and direct': General prompting advice from Anthropic, written about Claude: 'Claude responds well to clear, explicit instructions. Being specific about the output you want can help enhance results.' Use it only as the general principle of saying exactly what you want, not as a claim about GPT-5.5. (checked 2026-11-03)
- engine/ai-test.js runs, GPT-5.5 (OpenAI API), tests 'before', 'after', 'before-turn2', 'after-turn2' in tests.json: The real outputs: both prompts asked one question first. After a 38-word sample answer, the plain prompt returned two rewritten scripted answers (2,345 characters) with details not in the answer. The fixed prompt returned a two-line 'Worked / Improve' note and Question 2/5 (315 characters). (checked 2026-11-03)
Assumptions:
- Only GPT-5.5 (OpenAI API) was run. The two Claude entries in tests.json failed (claude -p was blocked in this session) and have null output, so the post must name GPT-5.5 (OpenAI API) only and never say 'tested in Claude' or 'ChatGPT app'.
- ai-test.js is single-turn, so the follow-up turn was simulated: the model's earlier question went in square brackets before the sample answer, in the same message as the original prompt. In a real chat the follow-up would be a second message. Say 'in my test', and don't claim a live multi-turn session.
- The job (marketing coordinator at a small outdoor gear shop) and the 38-word sample answer about a bakery's social media are fictional.
- The idea's premise that a plain prompt dumps ten questions at once did NOT hold: GPT-5.5 asked one question with both prompts. The post must not say the plain version dumped a list. The real difference is in the feedback step.
- The '35%' and '42%' in the plain output were the model's own example numbers, labeled 'For example'. Say it suggested numbers I don't have; don't say it claimed they were my results.
- Limitation: the fixed prompt's feedback is short and fairly generic ('add a specific result'). It doesn't coach delivery or tone, and it can't hear you speak.


