ChatGPT will praise a mediocre structure if you let it. That is not a bug in your prompting talent; it is the product's default posture. OpenAI's own Model Spec tells the assistant not to behave like a sycophant, yet in practice a free chat still drifts toward "nice try" unless you lock tone, exhibits, and scoring rules before the first turn.
This piece is the hands-on workflow. The sibling guide, what ChatGPT fixes and what it fakes, covers the judgment call. Here you get five copy-ready prompt sequences, the ChatGPT settings that keep numbers from being invented mid-case, and one honest note: Ned sells a scored alternative when you need grades that do not soften.
After you finish, you should be able to run a 35-minute ChatGPT block that generates a case, runs it interviewer-led, grades the structure, drills math, and stress-tests one PEI story without letting the model rewrite the data under you.
Lock the room before you paste a prompt
OpenAI's prompting guidance for ChatGPT is blunt: be clear, specific, and iterate. For case prep, that means three settings, not three adjectives.
| Setting | Where | Why it matters for cases |
|---|---|---|
| Custom instructions | Settings > Personalization | Persist "no praise, no invented exhibits" across chats (OpenAI Help) |
| Temporary chat | New chat > Temporary | Keeps a drill out of history and, if you choose non-personalized, off memory (temporary chat help) |
| Fresh chat per sequence | Each of the five blocks below | Stops yesterday's soft feedback from contaminating today's grade |
Paste this into custom instructions (or at the top of every temporary chat):
You are a consulting interviewer, not a tutor. Never praise. Never say "great," "solid," or "nice structure." Score only against the rubric I supply. If a number is not on the exhibit list, say "not in the data" and wait. Never invent revenue, margins, volumes, or market figures mid-case. Prefer "I don't know" to a confident guess.
That last line tracks OpenAI's public argument that models are trained to guess when scoreboards punish abstention (Why language models hallucinate). Your job is to reverse the incentive in the prompt.
Ned's rule. If the model can change the exhibit after you ask for a number, you are not practicing a case. You are practicing negotiation with a calculator that wants you to feel smart.
Sequence 1: Generate a case prompt (and freeze the exhibits)
Use a temporary chat. Goal: one interviewer-led prompt plus a closed exhibit table. Do not start answering yet.
Create one original interviewer-led profitability case for a fictional B2B client.
Constraints:
- Industry: industrial adhesives for packaging OEMs
- Client name: Northridge Bonding (fictional)
- One sentence situation, then a closed exhibit table with exactly these rows: 2024 revenue ($m), 2025 revenue ($m), gross margin %, SG&A ($m), volume (k tons), price ($/ton)
- Invent plausible but messy numbers once. After you print the table, those numbers are frozen for the whole conversation.
- Do not reveal a recommended structure or solution.
- End with: "Say ready when you want the opener."
Worked check you can reproduce: suppose the table prints revenue 184 and 171, gross margin 38% then 34%, SG&A 41 then 44, volume 92 then 88, price 2,000 then 1,943. Gross profit falls from about 70 to 58; operating profit after SG&A falls from about 29 to 14. If ChatGPT later "helps" by changing 2025 volume to 95 without saying so, kill the chat and regenerate. Invented data is the failure mode OpenAI documents; your exhibit list is the guardrail.
Sequence 2: Run the interviewer-led case with strict rules
New temporary chat. Paste the frozen case from Sequence 1, then this operator prompt:
You are the interviewer. Rules:
1. You drive the case. Ask one question at a time.
2. Exhibits are frozen. If I ask for a number not listed, reply only: "Not in the data. What else do you need?"
3. After each of my answers, reply in this exact format:
Pushback: (one skeptical sentence)
Next ask: (one question)
4. No praise. No hints that reveal the answer.
5. After my final recommendation, stop. Do not grade yet.
Begin with a 20-second verbal opener, then wait.
Run it out loud if you can. Typed turns are fine for volume; the interview room is spoken. I stop listening in real interviews when a candidate needs the interviewer to rescue the structure. This prompt forces the rescue attempt to fail.
If the model breaks format and starts coaching ("have you considered volume?"), reply once with: "Rules still apply. Pushback + Next ask only." If it breaks again, new temporary chat. Consistency beats a longer prompt stack; OpenAI's own API prompt guidance favors clear outcomes and stopping conditions over endless ALWAYS/NEVER clauses.
Sequence 3: Critique a structure against a rubric
New chat. Paste only your structure (bullets), not the full transcript.
Grade this case structure against the rubric below. Output only a table with columns Criterion, Score (0-2), Evidence (quote my words), Fix (one sentence). No encouragement.
Rubric:
- MECE: branches do not overlap; no silent catch-all
- Hypothesis-led: opens with a testable claim, not a laundry list
- Driver link: each branch could move profit for THIS client
- Prioritization: states which branch first and why
- Ask readiness: ends with the first data ask
Structure:
[paste]
A 0–2 scale keeps the model from handing out 8/10s. If every criterion lands at 2, your structure is too safe or the model is cheating; regenerate with "re-score; assume at least one 0 or 1."
Save the table. Tomorrow, rewrite only the weakest criterion and re-grade in a fresh chat. That is how structure improves without collecting five unread ChatGPT essays.
Sequence 4: Drill mental math (closed answer key)
New temporary chat. Math is where ChatGPT is useful for volume and dangerous for trust. Force an answer key up front.
Give me 8 mental-math items for consulting cases. Mix: % change, contribution margin, breakeven units, and simple CAGR.
For each item print:
Q: ...
Key: (the exact numeric answer and one-line method)
Then wait. I will answer one at a time.
After each of my answers, reply only:
Correct/Incorrect | My key | Error type (units / arithmetic / formula / none) | Next Q
No coaching paragraphs. Time expectation: ~45 seconds per item.
Run eight. Log your error types on paper. That log is the only progress tracker ChatGPT will not invent for you.
Sequence 5: Rehearse a PEI probe chain
Behavioral fails when the story is polished and the role is vague. Force a probe chain, not a rewrite.
You are running a Personal Experience Interview. I will tell one leadership story in under 90 seconds.
Rules:
- Ask exactly four follow-ups, one at a time, in this order:
1) What was YOUR specific decision, not the team's?
2) Who resisted, and what did they say?
3) What metric moved, by how much, over what period?
4) What would you do differently with the same facts?
- After my fourth answer, grade only: Role clarity (0-2), Resistance (0-2), Measurable outcome (0-2), Learning (0-2). One evidence quote each. No praise.
Say "Go" and wait for my story.
If the model starts rewriting your anecdote into a better speech, stop it. You need pressure on the weak criterion, not a ghostwriter.
A 35-minute block that actually sticks
| Minutes | Sequence | Output you keep |
|---|---|---|
| 0–5 | Settings + Sequence 1 | Frozen exhibit table |
| 5–20 | Sequence 2 | Spoken recommendation |
| 20–25 | Sequence 3 | Rubric table |
| 25–32 | Sequence 4 | Error-type tally |
| 32–35 | Sequence 5 | Four PEI scores |
Do this three weekdays a week. On one day, replace Sequence 2 with a scored live case so a second system grades you without negotiating.
When ChatGPT is enough, and when it is not
ChatGPT is a free drill operator if you freeze exhibits and ban praise. It is a weak grader for tracking weak spots across weeks, and it will not show a chart exhibit the way a purpose-built case tool does. I built CoachNed for that second job: everything is open for seven days, no card; then $120 for a recruiting season or $49 a month. Use /start for a five-minute typed rep with instant scores, /interview for the ~18-minute voice case, /drills/math when Sequence 4 gets soft, and /behavioral when you want four PEI follow-ups that do not rewrite your story.
CoachNed is not affiliated with McKinsey, BCG, Bain, Deloitte, or OpenAI.
Run this today
- Set the custom instruction block above.
- Complete Sequences 1–3 once on a temporary chat.
- Then take one scored structure or math rep on CoachNed so you have a grade the model did not negotiate: /drills/structure or /drills/quick-math (free forever).
Frequently asked questions
Can ChatGPT replace a case interview partner?
For volume drills and cold structure practice, yes, if you freeze exhibits and ban praise. For calibrated scores and chart exhibits, no. Pair daily ChatGPT sequences with one weekly scored case.
Should I use Voice Mode for cases?
Spoken practice matches the room better than typing. Keep the same strict operator prompt; Voice Mode does not remove sycophancy by itself.
Do custom GPTs help for case prep?
OpenAI's GPT FAQ notes that GPTs do not use your saved memory or custom instructions and each conversation starts fresh. Useful for a reusable interviewer script; still paste frozen exhibits every time.
Why does ChatGPT invent case numbers mid-interview?
Models are rewarded for confident answers. OpenAI's hallucination write-up argues that scoreboards that punish "I don't know" encourage guessing. Your exhibit list and "not in the data" rule reverse that for one chat.
Is CoachNed just ChatGPT with a wrapper?
No. ChatGPT is a general chatbot you steer with prompts. CoachNed is a scored practice product with fixed cases, drills, and debriefs. Use both: prompts for volume, Ned for grades that stick.
Sources
- Model Spec (2025-04-11) — assistant should not behave like a sycophant; push back when useful. Checked 2026-09-24.
- Why language models hallucinate — abstention vs guessing; why invented figures appear. Checked 2026-09-24.
- Prompt engineering best practices for ChatGPT — clarity, specificity, iterative refinement. Checked 2026-09-24.
- ChatGPT Custom Instructions — persistent behavior settings across chats. Checked 2026-09-24.
- Temporary chat in ChatGPT — history, memory, and personalization controls for drills. Checked 2026-09-24.
- GPTs in ChatGPT — GPTs ignore custom instructions and memory; fresh start each chat. Checked 2026-09-24.
- Prompt guidance (OpenAI API) — outcome-first prompts and stopping conditions. Checked 2026-09-24.