Prompting AI for L&D

A field guide for instructional designers with real deadlines. Written by someone who builds training with AI daily, not someone selling you a prompt pack.

Start here

You have three hours, not three weeks.

Every prompting guide assumes you're doing a tidy needs analysis first. You're not. The SME sent a 40-page SOP at 4pm, the launch is Friday, and "do ADDIE properly" is not on the menu. Good news: AI is most valuable exactly when the process is compressed, if you prompt it like a colleague instead of a search bar.

The one rule that matters

Context in, quality out. The model doesn't know your audience, your constraints, or what "good" looks like at your company. Every prompt that disappoints is usually missing one of those three. Everything else in this guide is that rule, applied.

Vending machine

"Write a training outline about data privacy."

Generic outline, corporate voice, ten modules nobody asked for.

Colleague

"Here's our SOP (pasted), the audience (new fraud agents, week 2), the deadline (Friday), and the behavior that must change (agents exporting data to spreadsheets). What's the smallest training that moves that behavior?"

Mindset

When you can't do ADDIE, do this instead

Compressed timeline triage, in one prompt each:

No time for analysis? Have AI interview you for five minutes: "Ask me one question at a time until you can tell me whether this is a training problem." You'll surface more than most kickoff meetings.
No time for design? Ask for three outline options at different sizes (15 min / 45 min / half day) with a recommendation. Pick, don't create.
No time for evaluation planning? One line: "What single question would tell us in 30 days whether this worked?" Put that in the plan before you build anything.

The phases don't disappear. They shrink to the size of the time you actually have — and a model makes shrunk phases survivable.

Interactive

Build the prompt, watch it improve

Toggle the ingredients and watch a mediocre request become one that gets a usable first draft. This is the same request, fed better.

Why each ingredient earns its place

Audience kills the generic corporate voice. "New fraud agents in week 2" writes differently than "employees."
Source material is the biggest single upgrade. Paste the ugly SOP, the SME transcript, the policy doc. The model grounds itself in your reality instead of inventing one.
Constraints (time, tool, length) prevent the ten-module monster nobody will build.
Output format means you get something you can paste into Rise or a storyboard, not an essay about learning.
Quality bar tells it what good looks like: "scenario-first, no bullet lectures, every section ends in a decision."

Five real Tuesdays

The scenario deck

Situations pulled from actual L&D life. Flip each between the prompt that wastes an hour and the one that works, then steal it.

Hard-won

Tricks of the trade

Make it interview you first. "Before you write anything, ask me questions one at a time until you have what you need." Five minutes of answering beats five rounds of bad drafts.

Paste the ugly source. Don't summarize the SOP for the model. The typos, the weird exceptions, the buried step 7b — that's where the real training need lives.

Ask for options, not answers. "Give me three approaches at different levels of effort, with what each trades away." You stay the designer; it does the drafting.

Make it disagree with you. "What's weak about this outline? What would a skeptical ops manager say?" Models flatter by default. Invite the red pen.

Give it your rubric. If you'd judge scenarios on realism, decision density, and voice, say so — then ask it to score its own draft against that rubric before showing you.

Stay in one conversation per project. Context accumulates. The tenth ask in a thread that has your SOP, audience, and constraints beats a fresh, amnesiac chat every time.

Keep a prompt library. When a prompt works, save it with a note about when to use it. Your library becomes your personal authoring tool inside a quarter.

The last 20% is yours. AI gets you a strong draft fast. The judgment — what to cut, what your org will actually tolerate, what the SME got wrong — is the job. That part was never the model's.

Built by Parker Lee · parkerlee.work · steal anything useful