Most candidates open Claude, type "run a case," and get a polite partner that never really pressures them. That wastes the only edge Claude has over a short chat: room for a whole casebook, your own transcripts, and a fixed interviewer persona that survives across sessions. After this piece you will know how to build that setup, paste five copy-ready prompt sequences, and tell when Claude is coaching you versus inventing numbers.
What Claude actually gives you for case prep
As of September 2026, Anthropic's paid Claude plans advertise context windows from about 200K tokens (roughly 500 pages) up to 1M tokens depending on the model. That is not a marketing flourish; it is why you can drop an MBA casebook PDF or three recorded debriefs into one workspace and still ask follow-ups that reference page 47.
Projects turn that window into a workspace: upload files once, set project instructions once, then open new chats that share the same knowledge base. Free accounts can create up to five projects; paid plans remove that cap and unlock RAG when project knowledge approaches the context limit. Artifacts sit beside the chat as editable documents or tools, which is where frameworks, score sheets, and sensitivity tables belong instead of drowning in scrollback.
Usage on every plan resets on a rolling five-hour window with no fixed message quota. Free covers light daily work; Pro is priced as at least 5x Free usage per window. Plan around that if you intend to paste long PDFs midweek.
| Claude feature | Case-prep use | Free tier (as of 2026-09-24) |
|---|---|---|
| Long context | Full casebook or multi-page transcript in one chat | Up to 1M tokens by model; Sonnet/Haiku on Free |
| Projects | Persistent interviewer persona + knowledge base | Up to 5 projects; no RAG expansion |
| Artifacts | Framework cards, math worksheets, debrief sheets | Available on Free |
| Usage pool | Caps how many long pastes you can run | Resets every ~5 hours; no fixed message count |
The thesis in one line: Claude is a memory machine for case prep, not a free partner who grades like a firm. Use the memory. Do not trust the grade.
Build one Project before you run a single prompt
Create a Project named something dull: "Case interviewer — HarborFresh practice." Upload only materials you have rights to use: your own notes, school casebooks your program licenses, and transcripts you typed yourself. Do not dump copyrighted commercial packs you never bought.
Paste this into project instructions (edit the firm flavor once):
You are a first-round interviewer at a global strategy firm. Stay in character.
Ask one question at a time. Do not praise. Do not invent exhibits or numbers.
If I ask for facts not in the case text or uploaded files, say "not provided" and force an assumption I must label.
After each of my answers, score Structure, Math, Insight, Communication from 1–5 with one sentence of evidence each.
Never soften a 2 into a 3 to be encouraging.
That instruction block is the product. Without it, Claude reverts to helpful tutor mode. Anthropic's own research on sycophancy in RLHF models is the reason you write the anti-praise rules yourself: preference-trained assistants still lean toward agreeing with the user when the prompt does not forbid it. Claude's constitution explicitly wants less sycophancy; your instructions still have to enforce it for interview practice.
Ask Claude to open an Artifact titled "Session scorecard" with columns for turn, your answer summary, and the four scores. Update that Artifact after every case so the Project accumulates a visible trail of weak spots.
Five copy-ready prompt sequences
Run these inside the Project. Paste the sequence as one message, then answer turn by turn. Do not ask Claude to "be encouraging."
1. Persistent interviewer cold open
Using only the Project instructions and any uploaded case text, open a 25-minute profitability case.
Give me the client, the ask, and one exhibit of numbers you invent only if I have not uploaded a case; label every invented figure as ASSUMED.
Then wait. Do not advance until I structure.
2. Full-casebook drill (one case at a time)
I uploaded a casebook PDF. Pick case N (I will name the page or title).
Summarize only: client, objective, available data. Quote no more than three sentences.
Run the case. When I finish, list every place I used a number that was not in the case text.
3. Transcript critique at interviewer depth
Below is my full transcript from yesterday's practice (unedited). Critique as a partner, not a coach.
Mark: (a) first 60 seconds, (b) structure MECE failures, (c) math steps with wrong units, (d) synthesis that did not lead with the answer.
Quote my exact phrases when you criticize. Do not rewrite the whole case for me until I ask.
[paste transcript]
4. Sensitivity chain with forced verification
Revenue $480m, COGS 62%, SG&A $95m, D&A $18m. Walk me through operating margin if COGS rises 4 points and SG&A falls 8%.
After each arithmetic step, pause and ask me to restate the intermediate number before you continue.
If I am wrong, correct once, then make me recompute. Do not invent a "market reaction" story.
5. Anti-sycophancy lock for PEI / fit stories
I will paste a 400-word leadership story. Grade only: my specific role, resistance I faced, measurable outcome.
Ask one partner-style follow-up that challenges the weakest claim.
If the story is vague, say so in the first sentence. No softening language.
[paste story]
These five cover the jobs Claude's context is built for. They are not a substitute for speaking aloud under a clock.
Worked example: HarborFresh margin pressure
HarborFresh sells meal kits. Baseline year: revenue $480m, COGS 62% of revenue, SG&A $95m, D&A $18m. You want operating margin after COGS rises four percentage points and SG&A falls 8%.
| Line | Baseline | After shock | Notes |
|---|---|---|---|
| Revenue | $480m | $480m | Hold volume for the drill |
| COGS | 0.62 × 480 = $297.6m | 0.66 × 480 = $316.8m | +4 points |
| Gross profit | $182.4m | $163.2m | |
| SG&A | $95.0m | $87.4m | −8% |
| D&A | $18.0m | $18.0m | Unchanged |
| Operating profit | $69.4m | $57.8m | |
| Operating margin | 14.5% | 12.0% | Drop of 2.5 points |
I stop listening when a candidate says "margin falls a bit" without stating 14.5% to 12.0%. Claude will happily narrate the story if you let it; Prompt 4 above forces the intermediate checks. Reproduce the table yourself before you trust any model output.
Ned's rule. Treat every number Claude speaks as a draft until you have recomputed it on paper. Long context raises recall; it does not raise truthfulness about arithmetic you never verified.
Guardrails against sycophancy and invented numbers
Three rules belong in every Project instruction and every hard session:
- No praise. "Good structure" is banned unless tied to a concrete MECE check. Praise is how models leak points you did not earn.
- No orphan exhibits. If a figure is not in the uploaded case or labeled ASSUMED by you, Claude must refuse to use it.
- Score with evidence. A 1–5 without a quoted phrase from your answer is invalid; make Claude regenerate.
Anthropic documented that RLHF-trained assistants can prefer user-aligned answers over truthful ones. That paper is from 2023; the failure mode is still the one that ruins case practice in 2026. Your job is to make agreement expensive in the prompt.
When Claude invents a "customer survey that shows 18% churn," stop the case. Paste: "Delete that exhibit. Only use uploaded data. Resume from my last structure." Then rebuild the Artifact scorecard so the invented path does not pollute the Project memory.
Where Claude still fails as an interviewer
Claude will not clock you. It will not cut you off at minute twelve because the room is behind. It will not give you the cold face a partner gives when your synthesis buries the answer. Voice delivery, interruption handling, and chart reading under silence still need a human or a product built for those constraints.
It also does not own a firm scoring sheet. The 1–5 rubrics above are useful heuristics; they are not McKinsey, BCG, or Bain calibration. CoachNed exists because I got tired of candidates mistaking fluent AI chat for interview readiness: everything is open for seven days with no card, then $120 for a recruiting season or $49 a month, with scored voice cases and drills that stay consistent across reps. Use Claude for the long-memory critique; use a scored product for the weekly reality check.
For a different stack aimed at high-volume short prompts, see the sibling ChatGPT workflow guide when you need volume instead of depth.
Practice this today
- Create one Claude Project. Paste the interviewer instruction block. Upload one case you already own.
- Run Prompt 1, then Prompt 4 on HarborFresh or your uploaded case. Force the intermediate math checks.
- Take the same case on CoachNed's timed path so you hear what a scored debrief sounds like: start at /start for three typed turns, or /interview for the ~18-minute CopperCup voice case.
Frequently asked questions
Is Claude free for case interview prep?
Yes for light use. Anthropic's Free plan includes chat, Artifacts, and up to five Projects, with usage that resets on a rolling five-hour window and no fixed message count. Heavy PDF pastes will hit the Free ceiling faster than short chat drills.
Should I put an entire casebook in one Project?
Yes if you licensed it and the file fits. On Free, you lack RAG expansion when knowledge nears the context limit, so prefer one casebook per Project or strip to the cases you will actually run. On paid plans, RAG can stretch project knowledge further.
Can Claude replace a live case partner?
No. It replaces a patient reviewer of long transcripts and multi-step math. It does not replace timed speaking, interruption, or calibrated scoring. Keep one live or scored session per week.
How do I stop Claude from inventing exhibits?
Put "do not invent exhibits or numbers" in project instructions, label every assumed figure yourself, and abort any turn that introduces orphan data. Then ask Claude to list every number it used and map each to a source line.
What should I use Claude for versus ChatGPT?
Use Claude when the artifact is long: casebooks, full transcripts, PEI stories, chained sensitivities. Use ChatGPT-style short loops when you need many quick structure or brainstorming reps in a day. The tools overlap; the context budget is what differs.
Sources
- What are projects? (Claude Help Center) — Free cap of five Projects; RAG on paid plans. Checked 2026-09-24.
- How large is the context window on paid Claude plans? — 200K to 1M tokens by model. Checked 2026-09-24.
- What are artifacts and how do I use them? — Artifacts on Free and paid; side-panel editable outputs. Checked 2026-09-24.
- Plans and pricing (Anthropic) — Free vs Pro usage, Projects, five-hour usage windows. Checked 2026-09-24.
- Towards Understanding Sycophancy in Language Models (Anthropic) — RLHF models can prefer agreement over truth. Checked 2026-09-24.
- Claude's Constitution (Anthropic) — Product intent to avoid sycophancy; still requires user-side guardrails for practice. Checked 2026-09-24.