Personal experience

Over a series of workshops I brought together every function that touched the P&L — operations, planning, commercial, R&D, and supply chain. The goal was simple: build a true unit cost for every product line, not the one the financial model showed, but the one the business was actually living.

Every session was transcribed. Every cost dataset was loaded. Every stakeholder perspective was captured.

I fed all of it into AI — transcripts, videos, cost data — and built a visual value stream showing the cost at each process step and the total delivered cost per product line.

Two things emerged that nobody in the room had identified going in: R&D costs were being allocated to the wrong product lines with no weighting for actual intensity, and commercial incentives were structured in a way that actively rewarded margin-destructive behavior.

The leadership team's mental model — built on years of financial model outputs — was structurally wrong. The operational reality told a completely different story.

The workflow — 6 steps

Step 1 — Gather your inputs Collect everything that touches the problem — meeting transcripts, cost data, stakeholder notes, process documentation. Load them into NotebookLM as a single project. This becomes your queryable knowledge base for the engagement.

Step 2 — Run the RCA prompt in Claude Paste your key data and ask Claude to map bottlenecks, flag cost anomalies, and identify where assumptions don't match the numbers. The prompt is in section 02.

Step 3 — Build the value stream Ask Claude to structure a process-by-process cost view. Which steps carry the highest cost relative to output? Where does value leak? Where do the model assumptions diverge from operational reality?

⚡ The step that AI can't do for you

Between mapping the value stream and red-teaming your recommendation sits the most important moment in the entire process — and it's entirely human.

This is where your domain expertise determines the quality of everything that follows. The questions you ask Claude, the anomalies you tell it to look for, the business realities you use to pressure-test its output — all of that comes from you.

Claude can process data faster than any analyst. It cannot know that a 3% variance in R&D allocation is catastrophic in your client's margin structure, or that a commercial incentive that looks reasonable on paper creates perverse behavior in the field.

The operators and consultants who get the most from AI aren't the ones who prompt best. They're the ones who know their domain well enough to know what questions to ask — and when to push back on the answer.

Your expertise is the differentiator. Claude is the amplifier.

Step 4 — Challenge the output Push back on what Claude gives you. Use your operational knowledge to stress-test its conclusions. Ask it to go deeper on the anomalies that don't match your experience. This is the iteration loop that separates a sharp diagnostic from a generic summary.

Step 5 — Red-team your recommendation Before presenting your findings, run the red-team prompt. Ask Claude to challenge your conclusions with the same rigor you applied to the client's data. The prompt is in section 02.

Step 6 — Present with confidence You've stress-tested your own thinking before anyone else gets the chance. The board's hardest questions are ones you've already answered.

Tool ratings Ease of setup: 8/10 — NotebookLM upload takes 10 minutes, prompts are copy-paste ready Time saved: 4–6 hours of manual synthesis per engagement Cost: Free on all tools to start

Most analyses fail not because the data is wrong — but because the questions asked of the data are too narrow. These two prompts are designed to work in sequence. The first finds the real problem. The second makes sure your solution holds up under pressure.

PROMPT 1 — Root Cause Analysis Run this in Claude after loading your data into NotebookLM

You are a senior strategy consultant specializing in operational diagnostics. I am going to give you data, transcripts, and notes from a business performance analysis.

Your job:

1. SYMPTOM vs CAUSE Separate what the business thinks the problem is from what the data actually shows. Be explicit about the gap between the two.

2. VALUE STREAM MAP Break down cost or performance by process step. Flag which steps carry the highest cost relative to output or value delivered. Identify where financial model assumptions diverge from operational reality.

3. ANOMALIES Identify where the numbers don't match the narrative. What is being misattributed, misallocated, or ignored? Flag anything that looks structurally wrong — not just a bad quarter, but a bad assumption baked into how the business measures itself.

4. ROOT CAUSES Rank the top 3 root causes by impact. For each:

5. BLIND SPOTS What assumption does the leadership team appear to be making that the data does not support? What would they need to see to change their mental model?

Be direct. Prioritize ruthlessly. Flag where data is insufficient to draw conclusions. Do not summarize — diagnose.

[PASTE DATA / TRANSCRIPTS / NOTES HERE]

→ This is where your expertise matters most. Before running Prompt 2, review the output against your operational knowledge. Push back on anything that doesn't match the business reality you've observed. Refine the diagnosis until it reflects both what the data shows and what you know to be true from experience.

PROMPT 2 — Red-Team Your Recommendation Run this in Claude before any board or client presentation

You are a skeptical board member and former McKinsey senior partner. You have seen every consulting framework and sat through hundreds of recommendations. You are not easily impressed.

I am about to present the following recommendation:

[PASTE YOUR RECOMMENDATION HERE]

Your job — be ruthless:

1. STEEL MAN Make the strongest possible case for this recommendation. What is the most compelling version of this argument?

2. ATTACK Identify the top 3 weaknesses in my logic or evidence. Where is the reasoning thin? Where am I relying on assumption rather than proof?

3. HIDDEN ASSUMPTIONS What am I assuming that I have not explicitly proven? What would have to be true for this recommendation to be right — and how confident am I that those things are true?

4. KILL SHOT What single question from the board would be hardest for me to answer right now? Give me the exact question, word for word.

5. WHAT I'D NEED What additional data or analysis would make this recommendation bulletproof? Be specific — not "more data" but exactly what data and why it matters.

Do not be polite. I need to walk into this room prepared for the hardest version of this conversation.

→ After running this prompt, answer the kill shot question before you present. If you can't answer it confidently, you're not ready. Go back to the data.

Both prompts are there and sequenced correctly with the expert judgment bridge between them.

What matters this week

→ BCG: AI is now used in 70% of strategy engagements at top consulting firms

BCG's latest research shows AI-assisted analysis has become standard practice across leading strategy firms — used for everything from market sizing to operational diagnostics and root cause analysis.

Why it matters to you: your clients' other advisors are already running AI-assisted analysis on the same problems you're working on. The question is no longer whether to use AI in your engagements — it's whether your workflow is as sharp as the firm sitting across the table.

→ NotebookLM adds audio overviews for uploaded document sets

Google's NotebookLM now generates podcast-style audio summaries of any document set you upload — including multi-session workshop transcripts, research libraries, and data compilations.

Why it matters to you: if you're running multi-session diagnostics or stakeholder workshops, you can now get a synthesized audio brief of the full transcript set before diving into Claude for deep analysis. Useful for large engagements where reading every transcript isn't realistic.

→ Claude's extended thinking mode now available for complex reasoning tasks

Anthropic's latest Claude update includes an extended thinking mode that shows its full reasoning chain before delivering conclusions — particularly valuable for analytical tasks where auditing the logic matters as much as the output.

Why it matters to you: for root cause analysis you can now see exactly how Claude reached its conclusions — not just what it found. That transparency is critical when you're pressure-testing the output against your own operational knowledge before presenting to a client.

You have a decision you're sitting on right now. A recommendation you're about to present. A business problem that's been stubborn longer than it should be. A number in your P&L that doesn't match your intuition about where the real issue is.

This week, run the diagnostic.

Start here — pick your situation:

If you have a recommendation to present → go straight to Prompt 2. Paste your full recommendation — not a summary, the actual argument — and read the kill shot question carefully. If you can't answer it confidently, you're not ready to present. Go back to the data.

If you have a problem you can't diagnose → start with Prompt 1. Load everything you have — data, notes, email threads, whatever exists. Give Claude more context than you think it needs. The more operational detail you provide, the sharper the output.

If you have both → run them in sequence. Diagnose first, then red-team the solution. That's the full loop.

The one thing to remember between the two prompts

Review Claude's output against what you know to be true from your own experience in the business or with the client. Push back on anything that doesn't match operational reality. Refine the diagnosis before you build the recommendation.

That iteration — between what the data shows and what you know — is where the real insight lives. Claude surfaces the pattern. You determine whether it's real.

Going further

If you're running a multi-session diagnostic or stakeholder workshops, build your NotebookLM project first. Upload every transcript and data file before running either prompt. You'll have a queryable knowledge base that gets more valuable with every session you add.

Save both prompts to your Cursor prompt library from Issue #2. They're now one keystroke away for every engagement.