Most people talk to AI like they’re typing keywords into Google.
“best crm for clinics”
“fix my react bug”
“write a blog post about ai”
Then they get a thin, generic answer and decide the whole category is overhyped. You handed a fast junior a sticky note with three words on it. Of course it flopped.
Brief these tools the way you’d onboard a smart person who has never worked at your company. That habit usually beats chasing the newest model.
The search box habit
Search is great when you want ten blue links. AI is weak at that job and strong at a different one: doing work when you give it context.
If your prompt looks like a search query, you get search-query energy back. Short. Generic. Confident in places you won’t notice until later.
When you catch yourself typing like that, stop and rewrite it as a brief.
How to brief an AI like a new hire
Cover five things. Not a novel. Just enough that a real person wouldn’t have to guess.
1. What you’re building Not “a dashboard.” Something like: “a simple admin page that shows missed calls from last night and who owns follow-up.”
2. What done looks like Acceptance criteria in plain English. “It pulls from this API. It shows a table. Empty state if there’s nothing. Don’t invent columns.”
3. What not to touch Auth. Billing. Anything with patient data. The production database. Say it clearly. Models love to be helpful in the worst possible rooms.
4. Where the truth lives Repo paths. File names. Screenshots. A pasted error. A snippet of the SOP. If the answer is in a doc and you don’t attach it, you’re asking it to improvise.
5. The voice and constraints “Write like a normal person. No corporate fluff. Keep it under 800 words.” You often have to say that, or you get LinkedIn-flavored filler all over the draft.
A before and after that actually works
Bad: “write a voice agent prompt for healthcare”
Better: “Write a system prompt for an after-hours voice agent for a multi-location treatment admissions line. It should greet calmly, collect name, callback number, and reason for calling, check if they’re in crisis and escalate those immediately to a human path, and never give clinical advice. Keep responses short. Sound like a competent front desk, not a robot therapist. Output the prompt only.”
Same model. Very different usefulness.
Coding is the same game
When you ship code with an agent, don’t say “fix this.”
Say what broke, what you expected, what you saw, which file you already suspect, and what “fixed” means. Then ask for a small diff, not a rewrite of the app.
If it proposes a big rewrite, push back the way you would with a junior: “Smaller. Show me the patch.”
The better you brief, the less magical it feels
That’s the point.
It’s not a genie. It’s a power tool. Power tools still need you to measure twice.
A tiny template you can steal
Copy this and fill in the blanks:
- Goal:
- Done means:
- Do not:
- Context / files / screenshots:
- Output format:
- Tone:
If you only do that for a week, your “AI doesn’t work for me” complaints usually get quieter. Or you’ll discover the model was fine and the brief was the bug.
Bottom line
Stop yelling keywords into the box. Brief the tool like a new hire: goal, done means, do-nots, context, and format. Clear briefs beat clever prompts.
