Who this is for
Whoever writes status updates, estimate blurbs, appointment confirmations, or internal job summaries. Entering here first? Fine — but if you can't name the writing job in one sentence, start at 01.
Pick one draft type
About 5 min
One type only. Examples: customer status reply · estimate blurb from notes · internal job summary · appointment confirmation. Master one before you add a second.
Write the privacy rule
About 5 min
Write three lines and tape them where the office can see them:
| Line | Your rule |
|---|---|
| OK to paste | e.g. job type, town, schedule window, public price-book lines |
| Never paste | e.g. SSN, full card numbers, unrelated medical or legal details |
| Redact pattern | e.g. names → initials, addresses → town only |
Use the prompt pattern
About 10 min
Use this fixed skeleton every time. The first line is the whole lesson: the model drafts, it never decides facts. The printable kit bundles the privacy rule, this prompt, and the five-draft log:
↓ Download the drafting kit (PDF)
You draft only. Do not invent hours, prices, dates, or promises.
Source notes:
---
{paste allowed notes}
---
Job type: {status reply / estimate blurb / ...}
Constraints: {tone, length, must include / must not claim}
If source is missing or conflicting, say what's missing instead of guessing.
Output: draft only, ready for human edit.
Try this
Paste the skeleton into ChatGPT or Claude with one real job — a status reply you already sent this week, so you know the right answer.
You should see: a draft close to what you sent. Compare every number and date against your source. If the model invented anything, that's the lesson working — flag it, don't send it.
Paste a whole thread with customer secrets to “give it context.” Redact first: names → initials, addresses → town only.
Apply the review rule
About 5 min
Before send: check every number, date, promise, and name against the source. If anything is wrong → fix it or don't send. The model does not “almost count.” This rule is the product; the prompt is just the shortcut.
Send without reading. No send before the source check — not once, not “this one's simple.”
Do this week
- Produce five real drafts with the pattern.
- Log each: minutes saved (rough), corrections required, and — most valuable — one case where you refused to send because the source was thin.
Success check
- Privacy rule written and shared with anyone else using the prompt
- Prompt skeleton saved where the office can find it
- Five checked drafts completed
- At least one time you refused to send because source was thin
Common failure modes
| Failure | Fix |
|---|---|
| Pasting whole threads with secrets | Redact; shorten to needed facts |
| Sending without reading | Hard rule: no send before source check |
| Asking the model to set price | Price stays human / rate card |
| One-off chats, no saved prompt | Save the skeleton in a shared note |
When to stop / don't buy yet
Do not put this on a public chatbot. Do not buy “AI customer service” to replace a missing review habit. If drafts need heavy rewrite every time, fix note quality first (Module 01) before paying for automation.
That's a systems problem, not a prompt problem. Bring five corrected drafts to a free consult — they scope the fix.