# FAST-03 Projects: Prompt Engineering for the Workplace Provider: 7formations | Course: FAST-03 | Instructor: Wadood Chaudhary | One capstone and two mini projects Projects turn the five-part card, the pass method, and the quality gate into tools trainees use after the course. Scoring uses **Not yet / Meets / Strong**. Completion requires Meets on every capstone criterion; mini projects are formative unless the employer asks for scoring. One resubmission allowed. Fictional or sanitized material only; never real customer, patient, or employee data (G.L. c. 93H, 201 CMR 17.00). --- ## Capstone: My Team's Tested Prompt Library **Brief.** Build a prompt library of at least 8 tested prompts for recurring tasks in your team (or for Cape Harbor Inn's front desk if your role has no safe tasks). Every prompt uses the five-part card, is bound to a named source, carries a check step and gate rule, and has survived a partner cold test. Include a one-page "how we use prompts" guide for the team. **Deliverables** 1. Task inventory (half page): 10+ recurring tasks; which 8 you chose and why; which you excluded (and the reason: data, judgment, or no source). 2. Library: 8+ entries in the W4 format (task, owner, source, prompt, check and gate rule, strip list, last tested, cold-test result). 3. At least two multi-pass entries (pass prompts written out). 4. Gate log: 6+ real or practice outputs gated with one-sentence reasons; at least one revise and one reject. 5. Cold-test evidence: questions your testers asked and the instructions you added. 6. Team guide (one page): where the library lives, how to use it, strip rules, the three doors, how to report an overshare, review schedule. **Step plan** | Step | When | What you do | | --- | --- | --- | | 1 | After session 1 | Task inventory; draft 3 entries | | 2 | After session 2 | Add pass-method entries; run and log | | 3 | Session 3 | Gate drill; first cold tests | | 4 | Within 10 days | Complete 8 entries; second tester; team guide; submit | **Rubric** | Criterion | Not yet | Meets | Strong | | --- | --- | --- | --- | | Structure | Entries miss parts or combine jobs | All entries five-part, one job each | Entries include "do not use for" lines | | Source binding | Outside facts allowed | Every entry names and fences a source; Gaps rule | Forbidden additions named per task | | Revision | No multi-pass entries or passes change facts | Two pass entries with tone-only later passes | Diff check built into the pass prompts | | Gate | Reasons about style or missing log | 6+ gated outputs with fact or data reasons | Gate findings fed back into prompt fixes | | Reusability | No cold tests | Two cold tests; questions converted to instructions | Library retested in a second tool or after an update | | Data safety | Restricted data in any entry | Strip list on every entry; report path in the guide | Sanitized input slots designed into prompts | **Strong vs weak example (one entry)** | | Weak | Strong | | --- | --- | --- | | Task | "Customer emails" | "Reply to guest questions about a scheduled event" | | Prompt | "Answer this email nicely." | Role (front-desk associate → guest); fenced fact sheet; one job; under 100 words, warm; "not in source → we'll confirm + Gaps"; "no dates, prices, or policies not in the sheet" | | Strip list | none | names, contact details, children's details, health info, card digits, coworker remarks | | Gate rule | none | "Reject if any fact isn't in the sheet or if the input wasn't sanitized; complaints go to the manager" | | Tested | "works" | "2026-10-20, Copilot Chat; cold test by R. (asked 'which sheet?'; added 'current dated sheet from Events folder')" | --- ## Mini project 1: Thin-to-Strong Makeover Gallery **Brief.** Collect five thin prompts (your own past ones, sanitized, or invented examples). Rebuild each with the five-part card, run both versions, and present a before/after gallery showing what changed in the output. **Deliverables.** Five before/after pairs; for each: missing parts, output differences (facts invented, length, tone), and the one part that made the biggest difference. **Step plan.** (1) Collect; (2) score the thin versions on the rubric; (3) rebuild; (4) run both in new chats; (5) write the gallery (one slide or half page per pair). **Rubric** | Criterion | Not yet | Meets | Strong | | --- | --- | --- | --- | | Diagnosis | Missing parts not identified | Parts correctly identified for all five | Explains why each missing part caused a specific output problem | | Rebuild quality | Rebuilds still thin | All five rebuilt with five parts | Rebuilds work for a partner cold | | Evidence | No outputs compared | Outputs compared on facts, length, tone | Invented-fact counts tabulated before/after | **Strong vs weak example.** Weak: "The new prompt is better because it's longer." Strong: "The thin prompt produced 3 invented facts (rain date, shuttle, raffle); the rebuild produced 0. The Check line naming forbidden additions made the difference; removing it brought the rain date back." --- ## Mini project 2: Pass-Method Playbook for One Pattern **Brief.** Pick one workplace pattern (customer reply, internal summary, or read-ahead). Write a reusable four-pass playbook for it and demonstrate it on two different sanitized inputs. **Deliverables.** Four pass prompts with placeholders; an assumptions prompt; two worked runs with diff checks; a one-paragraph "when to start a new chat" rule for this pattern. **Step plan.** (1) Choose the pattern and reader; (2) write pass prompts; (3) run on input 1, diff check; (4) fix prompts; (5) run on input 2 in a new chat; (6) write the rule. **Rubric** | Criterion | Not yet | Meets | Strong | | --- | --- | --- | --- | | Pass discipline | Later passes change facts | Passes 3 and 4 change tone or length only | Diff check shows zero fact changes in both runs | | Reusability | Prompts tied to one input | Placeholders work on both inputs | A partner used the playbook on a third input | | Chaining judgment | No new-chat rule | Clear rule tied to job and source | Rule includes a drift symptom to watch for | **Strong vs weak example.** Weak: four prompts that each say "make it better." Strong: pass 3 reads "Rewrite for [reader]: [tone], under [N] words. Do not add, remove, or change any facts; answer the same questions," followed by a diff table showing no fact changes.