Generic interview prep burns hours on questions you’ll never be asked. A good prediction prompt narrows from “all behavioral questions” down to the 8 questions this panel, at this company, at this level is most likely to ask — so your prep time lands where it counts.
TL;DR
- Prediction prompts work only when you feed them specifics: the JD text, your real background, the company’s public hiring signal, and the level you’re interviewing for.
- For live company/interviewer research, pick a model with web access. As of June 2026, Gemini 3.1 Pro has the tightest native Google Search grounding, ChatGPT (GPT-5.5) browses well when you tell it to, and Claude Opus 4.7 is strongest at long-document reasoning but searches only when web tools are enabled.
- Use the 12 templates below as a sequence: predict the question set first, then segment by panel, then close coverage gaps before drafting answers.
- Always verify predicted-question lists against a real source (Glassdoor interview tab, Levels.fyi). AI confabulates company-specific detail it can’t actually see.
Who this is for
Candidates with limited prep time, career switchers entering an unfamiliar interview culture, and anyone tired of grinding 100 generic questions that the panel will never reach.
These prompts narrow the high-probability set. They do not replace depth on the topics you do still need to cover — predictions are a focus tool, not a shortcut around fundamentals.
Which AI model to run these on (June 2026)
The prediction quality depends almost entirely on whether the model can see current, company-specific signal. Match the model to the job:
| Model (plan) | Native web research | Best for | Notes |
|---|---|---|---|
| Gemini 3.1 Pro (Google AI Pro, $19.99/mo) | Native Google Search grounding | Live company/interviewer research; recent news, launches, Glassdoor pages | 1M-token context; cited answers |
| ChatGPT GPT-5.5 (Plus, $20/mo) | Browses on request | Drafting and role-play; quick rewrites | In-app context ~320 pages on Plus; full 1M only on $200 Pro |
| Claude Opus 4.7 (Max) / Sonnet 4.6 (Pro, $20/mo) | Web search when enabled | Reasoning over a long JD + your full resume in one pass | 1M-token context standard; Pro bundles Claude Code + Cowork |
Practical pattern: do the research templates (3, 4, 10) on Gemini for fresh grounding, then paste those findings into Claude or ChatGPT for the drafting templates so you stay in one consistent voice. Never let any model assert a specific interviewer detail it didn’t retrieve from a link you provided — that’s where hallucinated names and titles creep in.
Prompt anatomy: the six elements
Every prediction prompt should carry six things. Drop any one and the output drifts generic:
- Role — name the persona the AI plays: candidate, hiring manager, recruiter, bar raiser.
- Context — target role, company, level, your background.
- Goal — one deliverable: a question list, a script, an answer, a coverage plan.
- Constraints — count, banned phrases, must-include facts.
- Tone — confident / curious / measured; 2-3 anchors.
- Examples — paste 1-2 of your past answers so it learns your voice.
What these predict well
- Behavioral question sets calibrated to a specific JD
- Technical / case-style question shape by domain
- Interviewer-specific tendencies from public signal
- Final-round panel decomposition
- Recruiter-screen topic lists
12 copy-ready prompt templates
Paste your real details into each [placeholder] before running. Templates are written to be model-agnostic; for the research ones, enable web access first.
1. JD-driven question set
JD: [paste the full job description]. My background: [paste 5-8 bullet points or resume summary]. Predict the 10 most likely behavioural questions for this role. For each: why this question (which JD line triggered it), what story I'd tell, and the gap I should prepare for.Swap in: the JD text, your background
2. Level-calibrated questions
Role level: [IC4 / IC5 / Staff / EM / Director]. Predict 8 questions that probe THIS level: scope, ambiguity, cross-team influence, mentorship, prioritisation under constraint. Skip IC-coding questions for an EM role.Swap in: the level
3. Company-culture predictor (run with web access)
Company: [name]. Search their careers/hiring page, engineering or interview blog, and recent Glassdoor interview reviews. Identify their distinctive interview style (e.g., Amazon Leadership Principles, Stripe writing exercise, Netflix culture-deck values). Output 6 likely culture-driven questions and cite the source for each.Swap in: the company name
4. Interviewer public-signal read (run with web access)
Interviewer: [name]. Their LinkedIn summary: [paste]. Their public talks / posts: [paste links or text]. Using only this public-facing signal, predict their go-to questions and which of my stories would resonate. Do not infer private details; flag anything you're guessing.Swap in: the name, LinkedIn summary, public material
5. Final-round panel decomposition
Final round: 4 panels — Bar Raiser, Hiring Manager, Skip-level, Cross-functional. For each panel: 2 questions they're likely to ask and the one story I should lead with. Don't prep a single story for all four; segment by panel.6. Recruiter-screen probable list
For a 30-minute recruiter screen, predict 8 likely topics: brief background, motivation, comp expectations, notice period, location, remote / on-site, deal-breakers, next steps. Draft a 30-second answer for each.7. “Why are you leaving” prediction
I'm leaving [current company]. Predict 3 variants of this question they'll ask and 3 ways to answer without trash-talking. One answer should be the "boring true" version.Swap in: the current company
8. Behavioral buckets, ranked
Standard behavioural buckets: leadership, conflict, failure, growth, influence, decision under uncertainty, ambiguity, prioritisation. For my background [paste], rank these by likely interview frequency for this role. Prep the 4 most likely first.Swap in: your background
9. Reverse-engineer from a rejection
I was rejected from a similar role last quarter. Feedback: [paste]. Predict which question types likely surfaced the gap, and prep specifically for those next time.Swap in: the feedback
10. Last-week refresh (run with web access)
Interview tomorrow at [company]. Give me a 30-minute refresh: (1) the 5 most likely questions, (2) one phrase to anchor each opening, (3) the one thing NOT to do, (4) what to read about the company in the morning — check for any news or product launches in the past week.Swap in: the company
11. Cross-functional ambush questions
Cross-functional panel (e.g., a PM interviewing an engineer, or vice versa). Predict 5 ambush questions: how do you collaborate with X, how would you handle Y trade-off, what do you wish Z roles understood?12. Question coverage audit
I've prepped [number] stories: [list them]. Map each to interview question types. Identify gaps: which likely question would have no story to tell? Prep that gap before anything else.Swap in: the number and list of stories
Common mistakes
- No specific context. Without company, role, and level, the output is generic and useless.
- Asking the model to “be honest” with no real record. It confabulates a candidate it has never met.
- Reusing one answer across companies. Interviewers on a hiring committee compare notes; a recycled story stands out.
- No tone anchor. Answers read flat and interchangeable.
- Skipping fact-checks. AI invents dates, headcounts, and titles. Verify any company specific against the source.
- Treating the first draft as final. First drafts read AI-flavored. Rewrite in your own words.
- No peer or mentor review. A predicted question set is a hypothesis; someone in the role can confirm it.
How to push results further
- Paste real past answers to anchor the model to your voice, not a generic candidate’s.
- Ask the model to play interviewer first and grill you; weak answers expose themselves under pressure.
- Draft three versions; ship the third.
- Read every answer aloud — anything you stumble over needs rewriting.
- Keep a personal bank of phrasings that landed well in past rounds.
- Have a peer currently in the role pressure-test your predicted list.
- Time-box practice. Fatigue makes your delivery worse, not better.
FAQ
- Can recruiters tell when an answer was AI-written? Yes, when it’s generic and over-polished. Specific numbers, names, and a real story are the antidote — those are exactly what a model can’t invent for you.
- How much research is enough for one interview? Roughly 60-90 minutes for an important round. Beyond that, returns diminish fast; spend the rest practicing delivery.
- Which model should I use for live company research? As of June 2026, Gemini 3.1 Pro (Google AI Pro, $19.99/mo) has the cleanest native Google Search grounding for fresh company pages and reviews. ChatGPT and Claude can browse too, but you usually have to tell them to.
- When should I start salary research? Before you apply. Negotiation that starts only after the offer arrives is weak. Use Levels.fyi for tech comp and validate against 2-3 sources.
- Are Levels.fyi / Glassdoor numbers reliable? Use them as a baseline, with caveats — small samples skew. Cross-check 2-3 sources before quoting a range.
- How do I keep prep notes organized? One doc per company: research, questions to ask them, and which of your stories fit which predicted question.
- How often should I refresh research before a final round? Do a quick re-scan the day of — any company news or product launches in the past week can become a live question.