Chapter 7 — Reps

Eleven reps to turn “AI can summarize this” into “AI summarized this and I checked every claim.” They move from reading a long document at speed, to grounding a summary sentence by sentence, to catching fabricated numbers and citations, to verifying data. No coding — just you, a document, an AI, and the discipline to examine whether the answer is so. The keyboard is the gym. Do not read these. Run them.

Ground rules

• Type it yourself. Write your own prompts; do not copy the ones in the chapter word for word. The point is to learn to ask, not to paste.
• Run everything. Every prompt, every check, every hand-calculation. A rep you only thought about is a rep you did not do.
• Predict before you measure. Before you open a source to verify a summary, write down how many errors you expect to find. Then find them. The gap between your guess and the truth is the lesson.
• AI usage note (every rep). You will use AI on nearly every rep — good. End each one with one honest line: what you asked, what it got wrong, and what you verified against the source. You choose the tool. You own the verdict.
• Your default tool this week is BoodleBox. Sign in with your Concordia account at box.boodle.ai and use Attach Knowledge to upload the report so the AI answers grounded in it — a portable knowledge repository. (Fallback: any assistant that accepts file uploads — ChatGPT, Claude, Gemini — works the same way.)
• Keep a reps.txt. One short written reflection per rep, in a single document (Google Doc, Word, or plain text). This is graded thinking, not busywork.

Reps 1–3: Read at speed, then read with judgment

Rep 1 — Two summaries, one report

Open code/sample-report.txt. In BoodleBox, attach it as Knowledge (or upload it to any assistant), then summarize it twice: first with a lazy prompt (“summarize this”), then with a professional one that names the audience and the decision, fixes the shape (bullets + recommendations with dollar figures + one paragraph on what it did not measure), and demands a source quote for every number.

Reflection: Put the two summaries side by side. What did the lazy prompt leave out or flatten that the professional prompt caught? Which one would you have been comfortable forwarding — and why is that a slightly frightening answer?

Rep 2 — Extract, don’t just summarize

From the same report, run three extraction prompts, not summaries: (a) every recommendation with its cost or savings figure, marking “none stated” where there is none; (b) every question the report raises but does not answer; (c) the section on what the report did not measure.

Reflection: Extraction (b) and (c) are the asks most people never make. Which one surprised you most, and how would it change a decision made off this report?

Rep 3 — Make the AI leave the box empty

Ask the AI for a figure the report does not contain — for example, “What is the projected return on investment of the office consolidation?” (the report never states one). Note what it does. Then re-ask with the instruction “if the figure is not in the source, answer ‘not stated’ — do not estimate.”

Reflection: What did the AI do the first time? Did the second prompt actually stop it from inventing a number, or did it still hedge its way to one? What does that tell you about how much you have to fence in an AI around missing data?

Reps 4–6: Grounding — tie every claim to the source

Rep 4 — Ground a summary, sentence by sentence

Take the professional summary from Rep 1. Copy the worksheet in code/grounding-checklist.txt into your reps.txt. Number the summary’s sentences and, for each, hunt in code/sample-report.txt (Cmd-F / Ctrl-F) for the words that support it. Mark each supported / partial / unsupported and paste the source quote.

Reflection: How many of your own AI’s summary sentences turned out to be fully supported? Were the shaky ones numbers, names, or conclusions? Name the one you were most surprised to find you could not fully ground.

Rep 5 — Hunt the seeded fabrications

Open code/ai-summary-flawed.txt — a realistic summary of the same report with problems planted in it. Without peeking at the chapter’s taxonomy table first, find every claim you believe is wrong. Then go back and, for each, name its failure type (fabricated number, magnitude error, wrong identifier, fabricated citation, added/opposite claim, overstated causation).

Reflection: Predict-before-you-measure: how many did you expect to find, and how many did you actually find? Which fabrication was the hardest to catch, and what made it blend in?

Rep 6 — Make the AI produce its own receipts

Paste the flawed summary AND the source into an assistant and ask: “For each sentence of the summary, quote the exact sentence from the source that supports it, or write NONE FOUND.” Read what it returns.

Reflection: Did forcing the AI to cite its own receipts surface the fabrications faster than your manual hunt in Rep 5? Where did it write NONE FOUND — and did it ever claim support that, when you checked the quote, did not actually match? (It can. That is why you check the receipts too.)

Reps 7–9: Data, spreadsheets, and citations

Rep 7 — Verify a number the AI computed

Give an assistant code/quarterly-figures.csv and ask for the headcount-weighted average Tuesday occupancy across all four offices, with the formula. Then re-derive it yourself by hand: multiply each office’s Tuesday percentage by its headcount, sum, divide by the total headcount (which should be 1,240 — a built-in check).

Reflection: Did the AI’s number match your hand calculation? If it did, does that make you trust the next one? If it did not, where exactly did the formula go wrong — and would you have caught it if you had not re-derived it?

Rep 8 — Explain, then break, a formula

Ask the AI to write a spreadsheet formula for “attrition improvement from 2023 to 2024 for each office” using the CSV. Have it explain the formula in plain English. Then deliberately feed it a subtly wrong description (“2024 minus 2023” when improvement is a drop, so it should be 2023 minus 2024) and see whether it blindly follows you or catches the logic error.

Reflection: Who caught the mistake — you or the AI? What does that tell you about relying on an AI to notice when your own request is wrong?

Rep 9 — Catch a fabricated citation

Ask a research-capable assistant (or a deep-research mode) a factual question in your own field that would need a source — e.g., “What does the research say about the productivity effect of hybrid work?” Take every citation it gives and try to confirm it exists and actually says what was claimed.

Reflection: How many citations checked out cleanly? Did any point to a real source that did not actually support the claim, or to a source you could not find at all? Write the exact check you used to confirm or debunk one citation.

Reps 10–11: Decision support and knowledge management

Rep 10 — Thinking partner, not decider

Take a small real decision you actually face this week. Ask an AI to lay out options, pros, cons, risks, and “what would you need to know to choose.” Use it to widen your thinking — then make the decision yourself and write down the human context the AI never had.

Reflection: What did the AI surface that you had missed? And what did you know about the situation that no amount of prompting could have given it? That second list is exactly why the verdict stays with you.

Rep 11 — Ground your own knowledge base

Take three of your own real documents (a policy, a past report, a handbook page — nothing confidential; see Appendix C before you upload). Load them into a grounded tool — in BoodleBox, add them as Knowledge (the Knowledge area, or Attach Knowledge in a chat); NotebookLM works the same way — and ask three questions answerable only from those documents. Check that each answer cites the right passage.

Reflection: Did the grounded answers cite the actual source passage, or drift into the model’s general knowledge? Where would a grounded document tool genuinely save you or your team time — and where would you still have to verify?

Done? One Last Thing.

A miniature of Project 7, end to end. Produce one document with three parts:

1. The summary — an AI summary of code/sample-report.txt from a professional prompt, saved unedited, with a header naming the tool and model you used and “mid-2026.”
2. The grounding table — every summary sentence marked supported / partial / unsupported, with the source quote (or “none found”) and, for each problem, its failure type.
3. The corrected summary — a clean version you would actually put your name on, with every fabrication fixed and every number re-checked.

If you can do this for one report tonight, you can do Project 7. That is the whole job in miniature: the AI summarizes, you examine it to see if it is so, and you sign only what you checked.

Up next: Project 7 — Project P7: Summarize and Verify.