Chapter 6 — Reps

Eleven reps to turn “AI writes my emails” from a vague idea into a disciplined skill: draft with the tool, verify by hand, keep your voice, and never ship a commitment you can’t trace. The keyboard is the gym. Do not read these — do them, on real messages and the shipped starter files.

Ground rules

• Type it yourself. Write your prompts and your final drafts by hand. Do not copy-paste the AI’s output straight into a real send field — retype or rewrite it, because that is where your judgment re-engages.
• Run/try everything. Every rep involves actually opening a tool (ChatGPT, Claude, or Gemini — see Appendix A) and doing the thing. A rep you only imagined is a rep you did not do.
• Predict before you measure. Before the AI drafts, write one line: what do I expect it to get wrong? Then check. The gap between your prediction and the result is the learning.
• AI usage note (every rep). End each rep that uses AI with one honest line: what you asked, what it got wrong, and what you verified or fixed by hand. This is not busywork — it is the habit the whole course is building.
• Keep a reps.txt. One short written reflection per rep, in a doc you keep. Use fake or scrubbed data only — never paste real client, patient, or student information into a consumer tool (Appendix C).

Reps 1–3: Draft, verify, and keep your voice

Rep 1 — The draft-then-verify loop, once, all the way through

Pick a real (low-stakes, non-confidential) email you need to write this week. Run the full six-step loop from §6.2: frame → draft → read every word → verify the load-bearing parts → make it yours → send (or save). Write down each step’s output.

Reflection: Which step did you most want to skip? (For most people it’s “read every word.”) What would have gone out wrong if you had?

Rep 2 — One message, three voices

Take one short message and ask the tool to write it three ways: (a) formal and corporate, (b) warm and human, (c) matched to your own voice — paste two of your real past emails as the sample. Put the three side by side.

Reflection: What are the specific “tells” of the default AI voice (words, rhythm, openings) that are not you? List three you will cut on sight from now on.

Rep 3 — Predict the hallucination

Ask an AI to draft a reply that requires a fact you can check — a date, a policy, a person’s name, a number. Before you read it, predict: will it invent a specific here? Then check every specific against reality.

Reflection: Did it fabricate anything? Confident tone and correctness are unrelated — where did the draft sound most sure and turn out to need checking?

Reps 4–6: Triage the inbox

Rep 4 — Triage by hand first

Open code/inbox.txt. Before touching AI, sort all 17 messages yourself into four buckets: reply today, reply this week, FYI only, ignore/delete. Time yourself.

Reflection: Which three did you flag as most urgent, and why? What did you use to decide that isn’t visible in the subject line alone?

Rep 5 — Triage with AI, then compare

Now paste the subject lines and one-line previews from code/inbox.txt into an assistant and ask it to sort them the same four ways and flag anything urgent or risky. Compare its ranking to yours from Rep 4.

Reflection: Where did you and the AI disagree? Name at least one message where your knowledge of the world beat its ranking of the text.

Rep 6 — Find the trap

Look hard at messages #2 and #3 in code/inbox.txt — two notices about the same overdue invoice, from different senders, one demanding a wire to new bank details. Decide how you’d actually handle it, and write the out-of-band verification step.

Reflection: Why would an AI triage tool be more likely to rank the fraudulent one highly, not less? What is the exact human step that catches it?

Reps 7–9: The transcript and its action items

Rep 7 — Summarize the messy transcript

Feed code/meeting-transcript.txt to an assistant and ask for a structured summary in three buckets: Decisions, Discussion, Action Items (owner / task / due date). Save the raw AI output unedited.

Reflection: How clean does it look? “Clean” and “correct” are different — note one place where the tidy formatting could hide an error.

Rep 8 — Audit the extraction against the source

Go through the AI’s action items line by line against the actual transcript. Hunt specifically for: the mis-heard name (“Anderson”/“Henderson”), the corrected date (24th vs 25th), and the phantom task (the flyer that was put on hold). Mark each item keep / fix / kill with the reason.

Reflection: How many of the three traps did the AI fall into? For each, quote the transcript line that proves the AI wrong.

Rep 9 — Fill the owned-action-items table

Complete code/action-items-template.csv for the whole meeting. Every row must have a named owner, a real (or explicitly-vague) due date, and a source quote from the transcript. Mark the Verified By Hand column only for rows you personally traced.

Reflection: Which action items had no clear owner or no real date in the meeting? What did you write in those rows instead of inventing one?

Reps 10–11: Judgment and the hard message

Rep 10 — Draft the “no,” own the firmness

Use the tool to draft a polite decline or bad-news message (e.g., telling the Taco Cart the $340 quote is over the $300 cap, per inbox #12 and #16). Let it find the calm words — then you decide how much to concede and how firm to be.

Reflection: What did you change from the AI’s version, and why? Which part was the machine’s job (the words) and which was yours (the decision)?

Rep 11 — The two-minute trace, timed

Take any three action items from your Rep 9 table and time how long it takes to trace each back to its source line and confirm the owner. Total it.

Reflection: It’s usually only a couple of minutes for all three. Given how cheap that is, why do people skip it — and what does skipping it cost when notes go out wrong?

Done? One Last Thing.

A miniature of Project 6, end to end. In one folder, produce three short documents from the shipped starters:

1. inbox-triage.txt — your four-bucket triage of code/inbox.txt, with the two traps flagged and an out-of-band step noted.
2. meeting-notes.docx (or .txt) — clean notes from code/meeting-transcript.txt in Decisions / Discussion / Action Items shape, with the name, date, and phantom-task traps already fixed.
3. action-items.csv — your completed code/action-items-template.csv, every row traced to a source quote and marked Verified.

If you can do this for one inbox and one transcript tonight, you can do Project 6. That is the whole week in miniature: the AI drafts, you verify, you sign.

Up next: Project 6 — Project P6: Inbox & Meeting Makeover.