The dirty secret of “we’re an AI shop now” is that one person still writes every prompt like it’s a magic spell, and everyone else waits. That’s not use. That’s a new bottleneck with a better story.
I used to be that person. It felt useful. It also meant every request died in the queue. The models got faster. I did not.
The unlock was not a better model. It was treating briefing as a team skill, the same way you’d teach someone to write a ticket that a junior engineer could actually execute. Once three people can brief, the output is not 3x because of math. It’s more than that, because they stop waiting and start shipping the small stuff without a priest.
Briefing is the skill. Prompting is just typing.
People ask for prompt training. I give them brief training.
A brief has a goal, a definition of done, a do-not list, the source of truth, and the output shape. A prompt is how you wrap that for the tool in front of you. If they only learn cute phrases, they will chase prompt Twitter while their SOP still lives in someone’s head.
Make them practice on real work, not “write a poem about our brand.”
- turn this messy SOP page into a checklist an agent can follow
- draft the first pass of a follow-up SMS from this call note
- ask the coding agent for a small diff in this file
- extract fields from this transcript into our schema
Same muscles. Different surfaces. The muscle is specifying work.
Give them a template, then watch them skip steps
The template you want on a wiki, in the Notion, on a sticky, do not care:
- Goal
- Done means
- Do not
- Context (files, links, screenshots, the SOP)
- Output format
- Tone or audience
They will skip “do not” first. That’s how you get an agent emailing a customer or touching prod. They will skip context second. That’s how you get confident fiction.
Do not scold. I show the before and after on their actual task. Empty brief, mushy output. Filled brief, usable draft. Adults learn from the delta, not from a lunch-and-learn that could have been a PDF.
If someone says “AI doesn’t work,” you ask to see the brief. Half the time the brief is three words. That’s not a model problem. That’s a search-box habit. We already had that talk on this site.
Pair them like you would with a junior
Sit with them for two sessions. Let them drive. You only interrupt when they’re about to paste secrets, ask for a rewrite of the app, or skip the fence.
Then they do one task solo and you review the brief, not just the output. Reviewing the brief is how the skill sticks. If you only fix the output, they will keep bringing you raw ore forever.
Also teach them to reject work. “Smaller.” “Show me the patch.” “This invented a policy. Go back to the SOP.” Those sentences are part of the job. An agent is not a vending machine. If they treat it like one, they will eat whatever falls out.
Put rails in the tools, not only in training
Training decays. Rails don’t as much.
- a shared prompt library with the real briefs, versioned
- project instructions that include the fence
- tool wrappers that cannot send the keys
- schemas for anything that must land in a CRM
- examples of good output sitting next to the template
When the team grows, the library is the onboarding. New people should not have to reverse engineer your brain from Slack.
If you have a coding agent, a customer agent, and a writing agent, they can share the briefing template even if the tools differ. The template is the company. The tools are rental cars.
Measure use like an adult
Do not measure “prompts written.” you measure work that left the bottleneck:
- tickets triaged without me
- first drafts that needed a light edit, not a rewrite
- PRs that were small enough to review
- call notes that landed structured on the first try
If all AI usage still routes through one person, you have a hobby. If the team can brief and you only get pulled in for the weird cases, you have a system.
There will be messy weeks. Someone will publish a samey post. Someone will let the agent wander. That’s management, not a reason to take the keys back and become the wizard again.
Bottom line
Use multiplies when the brief is a shared language. Models are cheap compared to the time you spend being the only translator.
Teach the team to onboard an agent like a new hire. Give them a template, a fence, and a couple of paired hours. Then get out of the middle of every task.
You will still be useful. You’ll just be useful on the hard stuff, which is the job you actually wanted when you started playing with this gear in the first place.
