FAST-02 Generative AI for Everyday Work · 7formations · Instructor: Wadood Chaudhary · All examples fictional
## Why this session matters
Drafting is where most people start with AI, and where most AI mistakes reach customers. The fix is not a better tool. It is a habit: give the AI your facts, tell it to use only those facts, make it flag what's missing, and work in small rounds you can check. This session builds that habit on a real-looking task: a guest notice for Cape Harbor Inn (fictional, Chatham).
## Key concepts
### 1. How models work, and how they fail, in one paragraph
A generative model writes by predicting likely next words from patterns in its training. It has no built-in sense of what is true *for your company*: your prices, your policies, your dates. When you ask for a notice without supplying the facts, it fills the gaps with whatever is typical: "complimentary breakfast," "heated pool," "free parking." Those details sound right because they are common, which is exactly why they slip past a quick read. Longer inputs add a second risk: details in the middle of a long document can be dropped or blended with others.
### 2. Source binding
Source binding means three things in your prompt:
1. **Fence the source.** Paste the facts between clear markers (`---` or `` tags) so the model knows exactly what counts.
2. **Say "use only."** "Use only the policy sheet between the markers. Do not add facts from anywhere else."
3. **Name the missing-fact behavior.** "If a fact the reader needs is not in the sheet, write 'not in source' and list it under 'Gaps:'."
The Gaps line is the most useful line in the output. It turns what would have been invented details into a question list for a person who knows the answer.
### 3. Work in rounds: outline, fill, polish
| Round | You ask for | You check | Typical fix |
| --- | --- | --- | --- |
| 1. Outline | Headings only, for a named reader and length | Is this the right job? Anything missing or extra? | Remove a section the task didn't ask for |
| 2. Fill | Each heading filled from the source, with "not in source" and Gaps | Every claim traceable to a source line | Correct a date or fee; answer a gap |
| 3. Polish | Tone, length, reading level, "no new facts" | Nothing added, removed, or changed | Remove a "helpful" promise the polish added |
Small rounds are easier to check than one big draft, and a problem is cheapest to fix in the round where it starts.
### 4. Data rules and the shared-printer test
Drafting prompts often tempt people to paste "context": a booking list, a complaint thread, last month's emails. Ask what the draft actually needs. A general guest notice needs the policy sheet, not anyone's booking. Personal information under Massachusetts G.L. c. 93H and 201 CMR 17.00 (a name with an SSN, license or state ID, or financial account or card number), PHI, and confidential business data (pricing plans, bids, HR matters) never go in. Before you paste, ask: is this tool approved? Did I remove names and numbers? Would I leave this on the shared printer? Who checks the output before it goes out?
## Worked example: the pool-closure notice
**Thin prompt:** "Write a notice to May guests about the pool and check-in."
**Typical output problems:** "Our heated pool reopens May 15" (sheet says May 22); "early check-in is complimentary" (sheet says $30); "enjoy complimentary breakfast" (not in sheet).
**Grounded sequence:**
```text
Round 1: I'm guest services lead at Cape Harbor Inn. Using ONLY the policy
sheet between the markers, give me a heading-only outline for a 120-word
email to guests arriving May 1-21 about (1) the pool closure and (2) the
early check-in fee change. ---[sheet]---
Round 2: Fill each heading using only the sheet. If guests would need a
detail that isn't there, write "not in source" and list it under "Gaps:".
Round 3: Make it warmer and keep it under 120 words. Do not add, remove,
or change any facts.
```
**Gaps the grounded version raised:** where guests can swim instead (not in source); whether the $30 fee applies to bookings made before May 1 (not in source). Both go to the owner. Neither should be guessed.
**Trace table (excerpt):**
| # | Claim | Sheet line |
| --- | --- | --- |
| 1 | Pool closed May 1-21 | "Pool: closed for resurfacing May 1-21" |
| 2 | Reopens May 22, weather permitting | "Reopens May 22 (weather permitting)" |
| 3 | Early check-in from 1:00 PM, $30 when available | "Early check-in from 1:00 PM: $30 ... subject to availability" |
## Checklist: grounded draft
- [ ] Source fenced between markers.
- [ ] "Use only" stated; outside facts forbidden.
- [ ] "Not in source" and a Gaps line required.
- [ ] Outline confirmed before filling.
- [ ] Every sentence traced to a source line.
- [ ] Polish round told "no new facts," and checked.
- [ ] Nothing restricted in the prompt.
## Do and don't
| Do | Don't |
| --- | --- |
| Paste the policy text the notice depends on | Rely on the model's "general knowledge" of your business |
| Ask for a Gaps line and route gaps to owners | Delete the Gaps line to make the draft look finished |
| Check the outline before asking for full text | Ask for a final draft in one shot |
| Recheck after every polish | Assume tone edits can't change facts |
| Paste only what the task needs | Paste booking lists or email threads "for context" |
## Try this at work
1. Pick one notice or letter you send repeatedly. Find (or write) the approved source text it depends on.
2. Run the three rounds in your approved tool with that source fenced.
3. Trace every sentence. Count how many claims you'd have sent unverified with a thin prompt.
4. Save the three prompts as a template for next time.
## Glossary
| Term | Meaning |
| --- | --- |
| Grounding / source binding | Telling the AI to write only from text you supply |
| Fence | Markers around the source so the model knows its boundaries |
| Gaps line | A list of facts the reader needs that the source doesn't contain |
| Round | One small step (outline, fill, or polish) checked before the next |
| Trace | Matching each claim in a draft to the source line it came from |
| Context window | The amount of text a model can take in at once; long inputs still lose detail |
## Key takeaways
- Models fill gaps with what's typical, not what's true for you.
- Fence the source, say "use only," require "not in source" and Gaps.
- Outline, fill, polish, and check each round.
*Tool features as of October 2026. Microsoft's documentation now uses "Microsoft Copilot" for Microsoft 365 Copilot; your screens may show either name.*