Design an AI Workflow That Sticks · Lesson 3 of 7

Lesson 3 — Designing prompts that survive real use

45 minutes. Move from "prompt that works once" to "prompt that holds up across hundreds of inputs."

---

The prompt you wrote in *AI in 60 Minutes* worked on the example you fed it. Real workflows feed it 50 different inputs a week. Some weird ones. Some adversarial. Some incomplete. The prompt has to hold up.

For any production workflow prompt, use this shape:

• Role — who the AI is acting as, with enough specificity that it knows the constraints (industry, size, voice).
• Task — what to produce, in one sentence.
• Constraints — the rules. Format, length, tone, what to never do, what to ask before doing.
• Output format — explicit structure. Headings, length per section, list vs paragraph.

Don't put the *input* in the prompt — put it in the user message, separate from the prompt. That way you can keep the prompt stable while inputs vary.

For any prompt, list the weird inputs it could receive:

• The input is too short
• The input is in a different language
• The input contains sensitive info (PII, financial data)
• The input contradicts itself
• The input is missing required fields

For each, write what AI should do. *Ask, don't guess. Flag, don't proceed.* These rules belong in your constraints section.

Once a prompt is in production, version it. Save the prompt with a date. When you tweak it, save the new version. When something breaks, you can compare.

Simplest version: a folder of `.md` files named like `customer-reply-v3.md`.

→ Lesson 4 — Connecting AI to email, calendar, docs

Previous: Lesson 2 — Picking the right hook · Next: Lesson 4 — Connecting AI to email, calendar, docs