How to write the perfect prompt (or close)

There is no perfect prompt. There are habits that keep the model from doing something technically correct and completely useless.

I build AI tools and sites for small businesses around DFW. The pattern I see over and over is not "wrong model." It is "vague ask." The model did what you typed. It did not do what you meant. If you are still picking tools, see How to choose which AI model is right for me first. Model choice and prompt habits work together.

The hose story (true, slightly painful)

I was editing a stock image for a project. The photo showed three disaster-relief machines under a portico (think portable awning), big machines with hoses.

I told the AI to remove two of the machines and leave only one.

It removed two machines. It left their hoses lying flat on the ground.

Technically obedient. Not what I wanted. The scene now looked like someone had ghosted the equipment and forgotten the plumbing.

Lesson: unless you are extremely specific, the model does exactly what you said, not what you pictured in your head. Funny. Useful. Slightly annoying when you are on a deadline.

A better ask would have sounded more like: "Remove the two machines on the left and right. Keep only the center machine. Also remove their hoses, cables, and any shadows or wet marks those machines left. The ground and portico should look clean and continuous, as if only one machine was ever there."

Same intent. Way more constraints. Way fewer surprise hoses.

Habits that beat "prompt magic"

1. Say what to remove AND what the scene should look like after

Deletion prompts fail when you only name the object. Name the aftermath.

• Bad: "Remove the logo."
• Better: "Remove the logo in the top right. Fill that area with the same brick texture and lighting as the wall around it. No blurry patch, no leftover outline."

2. Be specific about count, position, and keepers

"Leave one machine" is ambiguous if there are three. Say which one stays (left / center / right, or describe it). Same for people, columns in a spreadsheet, or paragraphs in a draft.

3. Add constraints the model will not invent for you

Tone, length, audience, format, banned phrases, must-include facts. Example for copy:

"Rewrite this homepage hero for a Texas HVAC company. Two sentences max. Plain spoken. No 'cutting-edge,' no 'revolutionize,' no em dashes. Mention same-day service once."

4. Do not assume the model knows your intent

It does not know your brand rules, your client's taste, or that "clean this up" means "delete the hoses too." If it matters, write it down.

5. Give a negative list and a positive list

• Negative: what must not appear (hoses, watermarks, stock-photo smiles, jargon).
• Positive: what must remain or appear (one machine, centered, dry concrete, natural shadow under the remaining unit).

6. Iterate in small moves

One big vague prompt invites one big vague mess. Change one thing per turn when you are close: "Now remove the left hose only" beats "make it perfect."

7. Paste an example when you can

"Match this tone" plus a short sample beats adjectives like "professional yet friendly," which every model has heard a million times.

A simple prompt checklist (steal this)

Before you hit send:

1. Goal: what success looks like in one sentence.
2. Keep: what must stay.
3. Remove / avoid: what must go or never appear.
4. Constraints: length, format, audience, tools, tone.
5. Context: who you are, who it is for, any facts that must be true.
6. Output shape: bullets, table, email, image edit notes, code only, etc.

If you cannot fill those six, the prompt is probably still mushy.

Etiquette that saves time (and money)

• One job per prompt when the tasks are different (research vs rewrite vs image edit).
• Tell it when to stop. "Ask me clarifying questions before drafting" is allowed and often smarter.
• Mark uncertainty. "If you are not sure about the address, say so instead of guessing."
• Do not paste secrets into consumer chats. Customer lists, passwords, health data, payroll. Use business plans and common sense. Talk to us if you want a sober look at where AI should and should not touch your ops.

When better prompts are not enough

Sometimes the model is wrong for the job. Image edit vs long legal rewrite vs live social research are different lanes. Sometimes you need a custom tool, a template library, or a human process, not another witty system prompt.

That is normal. Perfect prompts do not replace judgment.

How FlowLogic can help

At FlowLogic Solutions we help people get practical with AI: prompting habits, tool choice, and knowing when not to automate. If you want a clear map for your shop, start a conversation or browse Services.

Related reading: How to choose which AI model is right for me.