# FAST-03 Lab Guide: Prompt Engineering for the Workplace

Provider: 7formations  |  Course: FAST-03  |  Instructor: Wadood Chaudhary  |  Four labs (about 40 minutes each): Lab 1 in session 1, Lab 2 in session 2, Labs 3 and 4 in session 3
Audience: trainees (steps, worksheets) and instructors (setup, answer notes, timing). Instructor-only notes are marked **Instructor**.

All companies, people, emails, and data in this guide are fictional. Labs run in the employer's approved AI tool (Microsoft Copilot Chat, ChatGPT Business or Enterprise, Gemini for Google Workspace, or Claude for Work), in the 7formations sandbox, or on paper. Nothing in the labs connects to an employer's systems or uses real customer, patient, or employee data.

**A note on the five-part card.** The slides teach the card as **Role, Context, Task, Format, Check**. The course design's wording is "role, context, task, constraints, output." They are the same card: **Format** is the output specification and **Check** holds the constraints (what to flag, verify, or leave out). This guide uses the slide wording.

---

## 0. Lab environment and safety rules (read before Lab 1)

**Tools**

- Use only the AI tool your employer has approved, signed in with your work account. If none is approved, use the 7formations sandbox account (fictional data only, spending-capped, disabled at cohort end) or the paper versions.
- The instructor demonstrates in the employer's tool only. Menus change often; the instructor confirms current screens before each cohort.
- Paper fallback: every lab includes printed sample outputs. What is scored is the prompt and your judgment, not the tool.

**Data**

- Fictional data only. Under G.L. c. 93H and 201 CMR 17.00, a Massachusetts resident's name combined with a Social Security number, driver's license or state ID number, or financial account or card number is personal information. Never paste it into any AI tool.
- No PHI, no confidential business data, no children's details, no remarks about coworkers.
- Shared-printer test: if you wouldn't leave it on the office printer, don't paste it.
- If you paste something you shouldn't, stop and tell the instructor. Don't "ask the AI to forget"; it has already been sent.

**Fictional organization in this course**

- **Cape Harbor Inn**, a 40-room inn in Chatham, MA, with a restaurant (the Lantern Room) and a community events program. Every lab uses the inn's Beach Cleanup fact sheet.

**The fact sheet (all labs)**

```text
CAPE HARBOR INN - SPRING BEACH CLEANUP (fictional)
Date/time: Saturday, May 16, 9:00-11:30 AM
Place: Harding's Beach, Chatham. Meet at the east parking lot kiosk.
Provided by the inn: gloves, bags, water.
Children: under 12 must be accompanied by an adult. No lifeguard on duty.
Sign-up: front desk or events@capeharborinn.example by Wednesday, May 13.
Thank-you: volunteers receive a 10% dining credit at the Lantern Room,
valid the same weekend (May 16-17).
Contact: Events Manager, front desk (508) 555-0142.
```

Not on the sheet (deliberately): a rain plan, parking fees, pets, prizes, transportation.

**Completion**

- Submit W1 to W4. Completion requires Meets on every rubric criterion (Appendix B) for the final deliverable (W4 plus the gate decision in W3). One resubmission within the course window.

---

## Lab 1 (Session 1): Rebuild a thin prompt

**Scenario.** The events manager asked a front-desk associate to "get the word out" about the cleanup. The associate typed: *"Write something nice about the cleanup so people come."* The result invented a rain date, a free shuttle, and a raffle. You'll rebuild the prompt so the output can be trusted.

**Goal.** Produce W1: a five-part prompt bound to the fact sheet, its output, and a complete/incomplete score against the five-part rubric.

**Data.** The fact sheet above; the thin prompt; printed output of the thin prompt (paper fallback).

**Steps**

1. **Run the thin prompt (5 min).** Run it as-is. Highlight every sentence with a fact not in the fact sheet.
2. **Diagnose (5 min).** Mark which of the five parts are missing. (Answer: everything except a vague task.)
3. **Rebuild (12 min).** Write each part on W1:
   - *Role:* who writes, who reads.
   - *Context:* the fact sheet between markers, with "use only."
   - *Task:* one job.
   - *Format:* words, paragraphs, tone, ending.
   - *Check:* "not in source," a Gaps line, and named forbidden additions (rain date, prices, prizes).
4. **Run and check (8 min).** Run the rebuilt prompt in a new chat. Compare every fact with the sheet. Read the Gaps line.
5. **Partner score (7 min).** Swap prompts. Score your partner's prompt on the rubric (complete/incomplete per part) and one improvement.
6. **Revise (3 min).** Make one change from your partner's note and rerun.

**Deliverable: W1 Rebuilt prompt**

| Part | My text | Complete? (Y/N) | Partner note |
| --- | --- | --- | --- |
| Role | | | |
| Context (fenced) | | | |
| Task (one job) | | | |
| Format | | | |
| Check | | | |
| Facts in output not in sheet (count) | | | |
| Gaps line content | | | |

**Completion criteria.** All five parts present; source fenced with "use only"; "not in source" or Gaps rule present; output contains zero facts not on the sheet (or each one is identified and removed); one partner-driven revision.

---

## Lab 2 (Session 2): Two passes on a customer reply

**Scenario.** A guest wrote to the front desk (sanitized version): *"We'd love to join the cleanup. What happens if it rains? And will someone watch my 8-year-old while I pick up trash? Also, is there parking?"* You'll draft the reply in passes.

**Goal.** Produce W2: pass-1 (headings), pass-2 (fill from source), and pass-3 (audience) prompts and outputs, plus a separate assumptions list, with no new facts after pass 2.

**Data.** Fact sheet; sanitized guest question above.

**Steps**

1. **Pass 1, headings (5 min).** Ask for headings only. Confirm all three questions are covered.
2. **Pass 2, fill (10 min).** Fill each heading using only the fact sheet. Rain and parking aren't in the sheet: the reply should say the inn will confirm, and Gaps should list them. The child answer must match the sheet exactly: adult required under 12, no lifeguard.
3. **Assumptions (5 min).** "List what you assumed and what's missing. Do not change the draft." Record the list.
4. **Pass 3, audience (8 min).** "Rewrite for a parent who wants reassurance: warm, plain, under 100 words. Do not add, remove, or change any facts; answer the same three questions."
5. **Diff check (7 min).** Compare pass 2 and pass 3 sentence by sentence. Did pass 3 add anything ("our friendly staff will keep an eye on the kids")? If yes, go back to pass-2 text and tighten the pass-3 prompt.
6. **Chain decision (5 min).** Write one line: would you continue this chat for a different guest's question tomorrow? (Expected: no; new job, new chat.)

**Deliverable: W2 Pass prompts**

| Pass | Prompt | Output OK? | New facts introduced? |
| --- | --- | --- | --- |
| 1 Headings | | | |
| 2 Fill | | | |
| Assumptions list | | n/a | n/a |
| 3 Audience | | | |
| 4 Polish (optional) | | | |

**Completion criteria.** Child-supervision answer matches the sheet exactly in every pass; rain and parking routed to Gaps, never invented; pass 3 changes tone or length only; assumptions listed separately; chain decision recorded.

---

## Lab 3 (Session 3): Strip the overshare and gate the output

**Scenario.** A front-desk colleague forwards this guest email and asks you to "just have AI answer it":

```text
From: Dana Whitcomb <dwhitcomb@example.net>
Hi, this is Dana Whitcomb at 14 Pleasant Bay Rd, Orleans, cell 508-555-0177.
My son Eli (7) has ADHD and wants to help at the cleanup. Can he come
without me if a friend's mom watches him? Also, the woman at your desk on
Tuesday was really rude, I think her name was Marcy. Please charge the
dining credit to my card ending 4417.
```

**Goal.** Produce W3: a strip list, a sanitized prompt, the output, and a gate decision (accept as draft / revise / reject) with a one-sentence reason.

**Data.** The email above (fictional), the fact sheet, and three printed sample outputs (A, B, C) for the gating drill.

**Steps**

1. **Strip (8 min).** List every item that must not go into a prompt. Expected: name, email, home address, phone, child's name, child's health condition, coworker's name and the complaint about her, card digits. Then decide what the task actually needs: "a guest asks whether a 7-year-old may attend with another adult; separately, a service complaint."
2. **Split the job (4 min).** The complaint isn't an AI drafting job; it goes to the manager. The card request is not a dining-credit process (credits are automatic). Note both.
3. **Sanitized prompt (8 min).** Write the five-part prompt for the attendance question only. Run it.
4. **Gate drill (12 min).** Gate the three printed outputs, then your own:
   - Output A: correct ("under 12 must be with an adult; another responsible adult is fine"). Watch for this: the sheet says "an adult," not "a parent." Is "another adult is fine" in the source? (It's an interpretation: revise to "accompanied by an adult" and route the question to the events manager.)
   - Output B: says "Eli is welcome!" (Reject: it came from a chat where the full email was pasted.)
   - Output C: answers parking instead (Reject: wrong job.)
5. **Report practice (3 min).** Write the two-line report you'd send if you had pasted the original email: what, when, which tool.

**Deliverable: W3 Overshare strip and gate decision**

| Item | Strip? | Why |
| --- | --- | --- |
| | | |

| Output | Door (accept / revise / reject) | One-sentence reason (facts or data, not style) |
| --- | --- | --- |
| A | | |
| B | | |
| C | | |
| Mine | | |

**Completion criteria.** All 8 strip items identified; complaint routed to a person; sanitized prompt contains none of the stripped items; every gate reason tied to facts or data.

---

## Lab 4 (Session 3): Prompt library swap test

**Scenario.** The inn's front desk wants a shared set of tested prompts so every associate answers events questions the same way.

**Goal.** Produce W4: a library of at least 5 tested prompts from your own work (sanitized) or the inn's tasks, each with metadata, and evidence that a partner ran them cold.

**Data.** Your prompts from Labs 1 to 3, your prompt-clinic prompt, and the fact sheet.

**Steps**

1. **Select (5 min).** Pick 5+ tasks: e.g., event announcement, guest reply, internal summary, read-ahead for the manager, overshare-safe complaint acknowledgment.
2. **Write entries (12 min).** Use the template below.
3. **Swap test (15 min).** Your partner runs two of your prompts cold without asking you anything, and writes down every question they wanted to ask.
4. **Fix and retest (8 min).** Turn each question into an instruction; rerun once.

**Deliverable: W4 Prompt library entry (one per prompt)**

| Field | Entry |
| --- | --- |
| Task name | |
| Owner | |
| Source to fence | |
| Prompt (five parts) | |
| Check step and gate rule | |
| Data to strip before use | |
| Last tested (date, tool) | |
| Partner cold-test result | |

**Completion criteria.** 5+ entries with all fields; at least two partner cold tests with questions recorded and fixed; every entry names its source and strip list.

---

## Appendix B: Rubric (Not yet / Meets / Strong)

| Criterion | Not yet | Meets | Strong |
| --- | --- | --- | --- |
| Structure | Parts missing or two jobs in one prompt | All five parts; one job per prompt | Parts are specific enough that a partner runs it cold |
| Source binding | No fence or outside facts allowed | Source fenced; "not in source" required; no outside facts | Named forbidden additions; Gaps routed to an owner |
| Revision | Later passes change facts or the job | Later passes change tone or length only | Diff check recorded; pass prompts reusable |
| Gate | Door missing or reason about style | Correct door with a one-sentence fact or data reason | Report step and prompt fix recorded |