Paste the job ad into a chat, tell the model it is the hiring manager, and ask it to put one question to you at a time and wait for your answer. That setup does most of the work. You get questions that match the actual role, follow-ups that poke at the weak half of your answer, and a critique whenever you ask for one.
The part worth thinking about first is where the rehearsal happens. Interview prep means typing your employment history, the name of the employer you’re leaving, the reason you’re leaving, and the salary you’re hoping for. In a cloud chatbot all of that lands in an account history tied to your email. On a model running on your own phone, there is nowhere for it to land.
What AI interview practice is good at #
It is genuinely useful for four things:
- Predicting the questions. Feed it the job ad and it will produce the fifteen or twenty questions that ad implies. Most of them will come up.
- Follow-ups. “You said you led the migration. What did you personally decide, and what did someone else decide?” That second question is where real interviews go, and it is the one people haven’t rehearsed.
- Structure. Ask it to score your answer against the STAR shape (situation, task, action, result) and it will tell you, accurately, that you spent forty seconds on background and never said what changed.
- Compression. “Cut this to ninety seconds and keep the number” is a prompt a small model handles well.
What it is bad at #
Judging how you sound. A text model reads what you typed. It cannot hear the filler words, the speed, or the sentence you trail off. Rehearse delivery out loud separately, ideally with the phone’s voice recorder so you can hear yourself.
Knowing anything current about the company. An offline model’s knowledge stops at its training cutoff, so it will not know about the reorg, the funding round or the new VP. Ask it about those and you’ll get invention. There’s more on that in what offline AI knows about current events.
Salary numbers. Any figure a language model gives you for “market rate for this role in this city” is a guess dressed as data. Get salary ranges from posted listings and your own network, then use the model only to rehearse saying the number without flinching.
A system prompt that makes it behave like an interviewer #
Models default to being helpful, which in an interview roleplay means answering for you. A short system prompt fixes that. Set it once for the chat:
You are the hiring manager for the role below. Interview me.
Rules:
- Ask ONE question, then stop and wait for my answer.
- Never answer your own question, and never write my answer for me.
- After each answer, give two lines of critique: what landed, what was
vague or missing. Then ask the follow-up a real interviewer would ask.
- Stay in character. No summaries, no encouragement, no bullet lists.
Job ad:
[paste it here]The “ask one question then stop” line matters more than it looks. Without it, small models fire off ten questions at once and the rehearsal turns into a reading exercise. “Never write my answer for me” is the other load-bearing rule, because a helpful model will happily draft a perfect STAR story about a job you never had, and you’ll absorb it.
If you set this as a per-chat system prompt rather than a first message, it keeps applying as the conversation gets long. More system prompt examples covers the general pattern.
Five rounds to run before the interview #
- Question prediction. Paste the ad. Ask for the twenty questions most likely to come up, sorted by likelihood, plus the three that would be hardest for someone with your background.
- Behavioral drills. Work through the behavioral questions one at a time in the roleplay. Ask for STAR scoring after each. Two or three rounds is usually enough for a story to tighten up.
- The ninety-second opener. “Tell me about yourself” is the question people most often waste. Write yours, then ask the model to cut it to ninety seconds spoken, keep one number, and end on why this role.
- The awkward ones. The gap in your dates. Why you’re leaving. The skill in the ad you don’t have. Rehearse these until the answer is short, true and not defensive. Small models are fine at this because the goal is repetition, not brilliance.
- Your questions for them. Ask it to generate questions you could ask that would only make sense if you’d read the job ad closely. Pick three.
Keep the details on your own phone #
A job search is one of the few times most people have a genuinely sensitive document on hand. Your resume names your current employer, your manager’s team, your salary band and often your home city. If you’re searching while employed, none of that belongs in a chat history that is tied to an account and retained on someone else’s terms.
Personal LLM runs open models such as Qwen 3.5 and Gemma 4 on the phone’s own chip, so the job ad, your resume text and your salary number stay in a chat on the device. There’s no account to attach them to and no server they get sent to. It’s free on iOS and Android, and a model download is the only thing that touches the network. The same logic applies to the document you write afterwards, which is covered in writing a cover letter without uploading your resume.
Which model and settings to use on a phone #
| Model | Size | Good for |
|---|---|---|
| Qwen 3.5 4B | 2.74 GB, needs 3+ GB RAM | The default. Fast enough that the roleplay keeps a rhythm |
| Qwen 3.5 9B | 5.68 GB, needs 6+ GB RAM | Sharper critique and better follow-ups, noticeably slower |
| Ministral 3 3B | 2.15 GB, 256k context | Long job ads plus your resume plus notes in one chat |
| Gemma 4 E2B | 2.04 GB, needs 3+ GB RAM | Older phones with less memory to spare |
Two settings are worth changing. Use a balanced temperature (around 0.7) for the roleplay itself so the interviewer varies its phrasing, and a precise one (around 0.3) when you switch to asking for critique, so the feedback stays literal instead of creative. If your app has a thinking or reasoning mode, leave it off for the back-and-forth. It slows every turn down for no gain in a roleplay, though it helps if you ask the model to plan a whole answer strategy. Settings that change answer quality goes through the rest.
One practical trick: a read-aloud feature turns the typed question into a spoken one. Have the question read to you, answer it out loud at full length the way you would in the room, then type the shortened version for critique. You get delivery practice and text feedback from the same session.
After the interview #
Open a new chat the moment you’re out and dump everything you remember: the questions, what you said, what the room reacted to, the names. Ask for a thank-you note draft from those notes, and keep the file. If a second round comes, that dump is the best prep material you will have, and it’s more accurate than anything you’ll recall a week later.
Frequently asked questions #
Can AI actually tell whether my answer is good? #
It can tell you reliably whether an answer has a structure, a specific action and a result, and whether it’s too long. It cannot tell you whether a particular hiring manager will like you, and it will be too generous if you ask “was that good?” Ask instead for what was missing, which produces a much more useful answer.
Can I practice out loud instead of typing? #
Not inside a text-only chat app, and most on-device AI apps have no microphone input at all. The workaround is to answer out loud first for delivery, then type a condensed version for the model to critique. If you want spoken practice with a transcript, use a separate voice recorder or a transcription app and paste the text in.
Will an interviewer know I prepared with AI? #
Not from the preparation itself, which is just rehearsal. It shows if you memorize a model-written answer and deliver it, because the language is smoother and less specific than the rest of your speech. Keep your own words and your own numbers, and use the model to cut and sharpen rather than to write.
Which AI model is best for interview prep on a phone? #
A 4B model handles roleplay and STAR critique well and answers fast enough to keep the conversation moving. A 9B model gives better follow-up questions if your phone has 6 GB of RAM or more and you don’t mind waiting a few extra seconds. Size guidance is in how much RAM you need to run a model.
Is it safe to paste a job description and my resume into a chatbot? #
A public job description is harmless. Your resume is not, especially while you’re still employed, because it identifies you, your employer and your salary history together. Either strip it to the parts you need for the exercise, or use an app that keeps the whole conversation on your device.