# FAST-02 Projects: Generative AI for Everyday Work

Provider: 7formations  |  Course: FAST-02  |  Instructor: Wadood Chaudhary  |  One capstone and two mini projects

Projects turn the four work patterns (grounded drafting, summarizing, extracting, reusable helpers) into products trainees keep using. 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: A Verified Work Pattern for My Team

**Brief.** Choose one recurring document task from your job (a notice, a summary, a comparison, or a status update). Build a complete, verified AI work pattern for it: a grounded prompt sequence, a saved helper, a verification routine, and before-and-after evidence that it helps. The result must be usable by a coworker without you.

**Deliverables**

1. Task profile (half page): what, how often, who reads it, sources used, data risks, and what is stripped before use.
2. Prompt sequence: the outline, fill, and polish prompts (or the extraction prompt) with the source-binding and missing-fact rules.
3. Helper specification (W4 format) and a screenshot or export of the configured helper in the approved tool (or the paper prompt card).
4. Verification routine: the five-fact check or cell audit adapted to the task, with one completed example.
5. Before-and-after log over at least two runs (W5 format) with a keep / adjust / stop decision.
6. A one-page "how to use this helper" sheet for coworkers.

**Step plan**

| Step | When | What you do |
| --- | --- | --- |
| 1 | After session 1 | Choose the task; write the task profile; list data to strip |
| 2 | After session 2 | Draft the prompt sequence; run once on sanitized or fictional input; five-fact check |
| 3 | After session 3 | If the task includes tables or numbers, add the extraction and spreadsheet step |
| 4 | Session 4 | Build the helper; partner cold test; first log entry |
| 5 | Within 10 days | Second run and log entry; write the coworker sheet; submit |

**Rubric**

| Criterion | Not yet | Meets | Strong |
| --- | --- | --- | --- |
| Grounding | Prompts allow outside facts or have no missing-fact rule | Source fenced; "use only"; "not in source" and Gaps line | Gaps routed to named owners; polish constrained to no new facts |
| Verification | No recorded check | Five facts (or every cell) checked, results recorded | Check built into the helper's output and the coworker sheet |
| Reusability | Partner couldn't run it | Partner ran the helper cold and got a usable result | Spec portable across two tools, or migration-ready |
| Data safety | Restricted data in prompts or helper files | Only approved, sanitized sources; strip list documented | Refusal rule in helper instructions tested |
| Value evidence | No log or no checking time | Two logged runs including checking; honest decision | Error trend tracked; process adjusted based on evidence |

**Example of strong vs weak work (fictional: weekly supplier comparison at Bayside Precision Parts)**

| Element | Weak | Strong |
| --- | --- | --- |
| Task profile | "Compare quotes faster with AI." | "Weekly, 3 to 5 resin quotes; reader: purchasing lead; source: supplier quotes with bank details removed; risk: invented lead times and silent unit conversions." |
| Prompt | "Make a table of these quotes and pick the best." | Fixed 8 columns with units, blanks for missing, source column, "no totals or recommendations." |
| Verification | "Checked it, looked fine." | Cell audit log: 24 cells, 1 wrong (min order 50 vs 500), 0 invented after rerun. |
| Log | "Saves lots of time." | "By hand 45 min; with helper 6 min + 25 min audit = 31 min; saving 14 min; keep, with audit narrowed to price/lead/min cells after 4 clean weeks." |

---

## Mini project 1: The Two-Reader Summary Pack

**Brief.** Pick a long, non-sensitive document relevant to your job (a public regulation, a vendor manual, an internal SOP without personal data). Produce summaries for two different readers and a verification log.

**Deliverables:** chunk plan; two summaries (each with named reader, type, length); five-fact check per summary; coverage note.

**Step plan:** choose document → chunk by section → summarize each chunk → combine for reader A and reader B → five-fact check each → coverage check against headings.

**Rubric**

| Criterion | Not yet | Meets | Strong |
| --- | --- | --- | --- |
| Reader fit | Same summary twice | Type, length, and emphasis differ by reader | Each ends with what the reader must do |
| Accuracy | Facts not checked | Five facts per summary checked; errors fixed | Includes a caught error and the prompt change that prevents it |
| Coverage | Not checked | Headings compared; omissions justified | Middle-section facts confirmed present |

**Strong vs weak.** Weak: two summaries that differ only in length. Strong: a supervisor brief leading with "what changed and what to enforce this week," and a new-hire list of exact rules, both checked against the same five facts.

---

## Mini project 2: The Extraction Audit Challenge

**Brief.** Build a comparison table from three or more fictional or public documents (for example, three public product spec sheets) and prove it has zero invented values.

**Deliverables:** column definitions with units; extraction prompt; audited table with marks; spreadsheet with formula-based comparison; two-sentence recommendation with unknowns stated.

**Step plan:** define columns → extract one document at a time → partner audit → fix and rerun → spreadsheet math → note.

**Rubric**

| Criterion | Not yet | Meets | Strong |
| --- | --- | --- | --- |
| Table design | Columns undefined or units mixed | Defined columns, units, source column | Handles conversions explicitly |
| Audit | Partial | Every cell marked; zero invented | Error types analyzed (misread, invented, reversed) |
| Analysis | AI did the math | Formulas in spreadsheet | Sensitivity to unknowns shown |

**Strong vs weak.** Weak: a table where a missing warranty period was filled with "1 year (standard)." Strong: the cell is blank, the note says "warranty not stated by Supplier B; request before ordering."