Three-word prompts.
The single most common mistake. If the answer is generic, your prompt was generic. Add the context and the format and run it again.
PLAYBOOK · 01 · HOW TO PROMPT · BY MARC KLEINMANN
Most people talk to an AI like a search box. Brief it like a new hire instead.
How to ask an AI for something and get a usable answer the first time. The four parts of a good prompt, the two calls you make on every one, and the before-and-after that shows the whole difference. Written for the operator tired of generic answers.
01 ·Context
The tool is fine. The prompt was three words.
Most people talk to an AI like a search box. They type three words, get a generic answer, and decide the tool is overhyped. The tool is fine. The prompt was three words.
A prompt is just an instruction. The quality of what comes back is set almost entirely by the quality of what you put in. Not by clever tricks or secret words. By telling it the four things any competent person would need if you handed them the same job: what you want, who it is for, what good looks like, and anything it cannot see.
You already know how to do this. You do it every time you delegate to a person. You would never tell an assistant "write the update" and walk away. You would say which client, what happened this week, how formal, and where the numbers are. A prompt is that same briefing, typed.
02 ·Architecture
A good prompt has four parts.
A good prompt has four parts. You do not need all four every time, but when an answer comes back wrong, it is almost always because one of these was missing.
Part 01 · The taskThe actual thing you want done, as a verb.
"Draft," "summarize," "compare," "rewrite." Not "thoughts on" or "help with."
Part 02 · The contextWhat it cannot see.
The audience, the situation, the constraint. "This goes to a client who is already annoyed." "We are a 12-person firm, not an enterprise."
Part 03 · The formatWhat good looks like.
"Five bullets, one sentence each." "A short email, no greeting." "A table with three columns." If you do not say, it picks for you, and it usually picks long.
Part 04 · The materialThe raw input, pasted in or pointed to.
The transcript, the draft, the numbers. An AI cannot read your mind or your other tabs. Give it the source.
03 ·Setup
Nothing to install. One upgrade worth making.
This works the same everywhere, so there is nothing to install.
In Claude, ChatGPT, or Gemini: type it in the message box.
The one upgrade worth making: if you keep typing the same context every time (who you are, how you write, what your business does), stop. Put it somewhere standing. In Claude that is a Project's instructions. In ChatGPT, custom instructions or a Project. In Gemini, a Gem. Then every prompt starts with that context already loaded, and your actual messages get short.
04 ·Decisions
Two calls you make on every prompt.
Decision 01Short prompt, or full briefing.
Use a short prompt for a throwaway question. Use the full four-part briefing the moment the output matters or anyone else will see it. The rule: the more the answer matters, the more you tell it up front.
Decision 02Fix the prompt, or fix the answer.
When the first answer is wrong, do not hand-edit it into shape. That teaches you nothing and you do it again tomorrow. Instead, tell it what was wrong in one line and let it redo the whole thing. "Too long, give me five bullets." "Too formal, this is a friendly client." You are training the instruction, not patching the output. This is exactly the habit that, repeated enough, becomes a skill (see Writing your first skill).
05 ·Runbook
The template, and the before-and-after.
Start from the blank template. Fill one part at a time. The task, the context, the format, the material.
prompt template · the four-part briefing
TASK: [the verb and the thing. "Draft a reply to this email."] CONTEXT: [what it cannot see. Who it is for, the situation, any constraint. "The sender is a client asking why we are behind. We are behind, but we have a fix."] FORMAT: [what good looks like. "A short, warm email. No greeting line. Three short paragraphs. End with a specific next step and a date."] MATERIAL: [paste the source here. The email, the draft, the notes, the numbers.]
Now the same idea, shown as a before and after. Same person, same tool, same job.
before and after
BEFORE (three words, generic answer):
write a client update
AFTER (four-part briefing, usable answer):
TASK: Draft my weekly status update to a client.
CONTEXT: The client is a 30-person law firm. We shipped their intake automation
Friday. I need one thing from them: sign-off on the new lead form copy. Friendly,
not formal. They are happy with us.
FORMAT: Short email. One line opening, what shipped, the one thing I need, what is
coming next week. One sentence per point.
MATERIAL: [paste the week's notes here]Read the two next to each other. Same person, same tool, same thirty seconds of typing. The only difference is that the second one briefed the AI like it briefs a new hire.
06 ·Gotchas
The four that catch everyone.
Hand-editing instead of re-prompting.
If you find yourself rewriting the output by hand every time, you are doing the AI's job. Tell it what was wrong in one line and let it redo it.
Forgetting the material.
The AI cannot see your other tabs, your files, or last week's conversation unless you bring it in. Paste the source.
Being polite instead of clear.
"Could you maybe help me think about possibly drafting something" wastes the briefing. It is not a person with feelings to manage. Say the task as a verb.
07 ·What's next
Once your prompts are good, they want to become skills.
Once your prompts are good, you will notice you type the same one over and over. That annoyance is the signal that it is time for a skill: the refined prompt, given a name and saved once, so you never type it again.