Chapter 2: Working with AI — CRAFT & Tools

🎯 Chapter Objectives

Chapter 1 gave you the what. Chapter 2 gives you the how. By the end of this chapter you will have a working prompt template you can reuse for almost any MPV task, and a clear decision framework for which AI tool to use when.

• Climb the prompting ladder: Question → Instruction → CRAFT Brief.
• Master CRAFT: Context · Role · Action · Format · Target audience.
• Apply CRAFT to four real MPV departments: Finance, Procurement, Factory Ops, Distribution.
• Learn four patterns that multiply CRAFT: Few-Shot, Chain-of-Thought, Constrained Output, Iterative Refinement.
• Map the 2026 AI tool landscape into four tiers: Foundation Models · Enterprise Platforms · Vertical AI · No-Code/Agents.
• Use a four-question decision framework to pick the right tool for each MPV task.
• Apply the "Rule of Thumb Before You Paste" — the single most important safety habit.

2.1 From Question to Structured Brief — The Prompting Ladder

Most users never get past Level 1. Their AI outputs disappoint, they blame the tool, they stop using it. Climbing one rung up the ladder produces 5× better results from the same model.

Level 1 — Question

Result: A generic CAPA template with placeholder text. You'll throw it away.

Level 2 — Instruction

Result: Usable first draft, but generic structure and may miss your QMS template fields.

Level 3 — CRAFT Brief

Result: A near-ready draft. You verify the facts (especially regulatory references), polish 10% of the language, and it's signable. Typical time saving: 60–90 minutes per CAPA.

2.2 The CRAFT Framework — Five Letters You'll Use Every Day

CRAFT is the simplest prompt structure that survives contact with real work. Memorise the letters. The order matters: each letter sets up what the AI needs to know before you tell it what to do.

C — Context: The Background Slide

For an MPV prompt, your default context paragraph mentions: company (disposable medical devices, ISO 13485:2016), location (Tam Diep IZ, Ninh Binh), today's situation, any specific data the AI needs (dates, numbers, lines, people). Two to four sentences. If you find yourself writing more than 5 sentences of context, you're probably trying to do too much in one prompt — split into two.

R — Role: Which Hat Do You Want AI to Wear?

Roles unlock different writing styles, structures, and assumptions inside the AI. Pick the role that matches your reader. Useful MPV roles:

• "Regulatory Affairs analyst familiar with ISO 13485" — for CAPA, deviation, audit response
• "Senior Vietnamese factory supervisor" — for SOPs, shift reports, operator-facing instructions
• "Customer service representative for a medical-device manufacturer" — for hospital/distributor replies
• "Cost accountant" — for finance memos, variance commentary, board summaries
• "Procurement officer used to RFQs from injection-molding suppliers" — for supplier comparisons, RFQ drafting

A — Action: The Specific Verb

The seven verbs that cover ~90% of MPV's daily AI use: Draft · Summarise · Classify · Translate · Compare · Recommend · Check. Use one per prompt. If you need two, run two prompts — the output of each will be sharper.

F — Format: The Container

Examples of useful format specifications:

• "A table with columns: Defect Type | Count | Line | Recommended Action"
• "Bullet list, each bullet under 15 words, max 7 bullets"
• "Formal Vietnamese email, 150–200 words, ending with a clear next-step request"
• "Markdown with section headings; each section 2–4 sentences"
• "JSON with fields: complaint_id, category, urgency_1to5, suggested_owner"

T — Target Audience: Who Reads This?

"For an ISO 13485 auditor in our 2027 surveillance audit" produces very different language than "for the Line 3 night shift operators reading on their phones during break". Specify it. The AI will calibrate every word choice.

2.3 Real CRAFT Prompts for MPV Departments

Four ready-to-use CRAFT prompts from four MPV departments. Copy, adapt, deploy. These are your starter prompt library for week one.

Finance — Board Commentary on Monthly Variance

Procurement — Resin RFQ Comparison

Factory Operations — SOP Starter from Engineer's Notes

Distribution — Delivery Delay Notification to a Hospital

2.4 Four Patterns That Multiply CRAFT

Once CRAFT is muscle memory, these four patterns lift quality further. Use them when the basic CRAFT prompt produces a 75% answer and you want a 95% answer.

1. Few-Shot Prompting — Give the AI Examples

Include 2–3 examples of the input→output pattern you want, then ask for the same pattern on a new input. Often the difference between an unusable and a usable output.

2. Chain-of-Thought — Ask AI to Think Step by Step

For analysis, comparison, or multi-step problems, add "think step by step before giving the final answer" or "show your reasoning". This forces the AI to lay out its logic — which both improves accuracy and makes errors visible to you.

3. Constrained Output — Lock the Shape

For automation pipelines, constrain to JSON / table / specific word count. Removes ambiguity downstream.

4. Iterative Refinement — Conversation, Not Single-Shot

The first response is rarely the best. Use follow-ups: "Make it 30% shorter." "Drop section 3." "Rewrite section 4 in a warmer tone." "Now translate to Vietnamese for the operator handout." Treat AI as a junior writer you're editing in real time — not a vending machine.

2.5 The Mental Map — Four Tiers of AI Tools

The 2026 AI tool landscape is overwhelming if you treat each product as unique. Group them into four tiers — then "which tool?" becomes "which tier?", a much smaller decision.

Quick Tool Guide — When to Reach for Which

2.6 Domestic Alternatives — Often Better for VN Workflows

Don't default to American tools just because they're famous. For Vietnamese-language workflows, data residency, and VND-priced contracts, the Vietnamese stack often wins. The leadership team will ask about these — know them.

2.7 The Decision Framework — Four Questions Before You Pick a Tool

Don't argue tool-by-tool. Run any AI-tool decision through these four questions, in this order. The answer falls out.

Q1 — Commodity or Differentiator?

Commodity capability (drafting an email, summarising a doc, translating): use whatever your team already uses. Don't waste a procurement cycle.
Differentiator capability (something that genuinely affects MPV's competitive position — a custom defect-detection model, a proprietary supplier-risk score): worth scoping carefully, possibly Phase 3 custom build.

Q2 — Data Sensitivity?

Q3 — How Fast Do You Need It?

Q4 — Vendor Lock-in & Trust

Three questions to ask every vendor before signing:

• Data residency: "Where is our data stored? Can you provide a Decree 13/2023-compliant DPA?"
• Exit: "If we leave you in 18 months, do we get our data and our model weights back, in a portable format?"
• AI Law compliance: "How will you support our compliance with Vietnam's AI Law and the 18-month healthcare grace period?"

2.8 The Rule of Thumb Before You Paste

The single most important safety habit in your AI life. Practice it until it's automatic.

📝 Chapter 2 Summary — What You Now Own