Use AI better
How to ask AI for better help without learning prompt engineering
Give useful context, say what outcome you want, and treat the first answer as the start of a conversation.
The short version
You usually do not need a magical prompt formula. Explain the situation, the outcome you want, the constraints that matter, and what the AI should not assume.
Optional explanationRead the full guide
Talk about the job, not the technology
“Write this better” gives the AI very little to work with. “This is a follow-up to a recruiter I spoke with last week. I want to sound interested without sounding pushy. Keep my tone warm and direct” gives it a real job.
Useful pattern
Here is the situation: [CONTEXT]. I want to accomplish: [OUTCOME]. Important constraints: [CONSTRAINTS]. Use only the facts I give you, and ask me a question if a missing detail would materially change the result.
Four things that improve almost any request
You do not need all four every time. Add the detail that would actually change the answer.
- Context: what is happening and who is involved?
- Outcome: what would a useful result let you do next?
- Constraints: time, budget, audience, tone, tools, age, location, or other limits.
- Boundaries: what should the AI not invent, assume, or decide for you?
Do not restart every time the first answer is imperfect
A conversation is often more effective than a giant one-shot prompt. Tell the AI exactly what missed the mark.
- “That sounds too corporate. Make it sound more like the notes I gave you.”
- “The plan is too busy. Keep one main activity per day.”
- “Separate confirmed facts from assumptions.”
- “Explain that again using an example from cooking.”
Examples beat adjectives
If “friendly,” “professional,” or “simple” is not producing what you mean, show the AI a short example or describe what you dislike. “No corporate filler, no exclamation marks, and keep it under six sentences” is much clearer than “make it better.”
Use a repair ladder when the first answer is bad
A weak first answer does not always mean the tool is bad. Often the fastest move is to tell the AI exactly which part failed and add only the missing information that would change the result.
- Name the problem: “This is too generic.”
- Add the missing constraint: “I only have 30 minutes and two kids with me.”
- Protect the facts: “Do not invent anything that is not in my notes.”
- Ask for choices: “Give me three versions with different tradeoffs.”
- Ask it to surface uncertainty: “What are you assuming that I did not tell you?”
Repair, do not restart
“The plan is too ambitious. Keep the same goal, but give me one main task, one optional task, and a realistic stopping point.”
The common mistake is chasing a perfect one-shot prompt
You do not need to squeeze every possible instruction into one giant message. A short first request followed by one or two corrections often produces something more useful and is much easier to remember.
Think conversation, not incantation.
Optional depthGo deeper when you want to
Better prompting is less about clever syntax and more about controlling ambiguity. The deeper method is to diagnose what the assistant would have to guess, then decide which guesses matter enough to remove.
Use an assumption audit before adding more prompt words
When an answer is strangely confident or generic, ask what the assistant had to assume: audience, timeline, priority, tone, budget, source authority, missing facts, or what ‘good’ means.
Then add only the assumptions that would materially change the result. This produces clearer requests without turning every prompt into a wall of instructions.
Assumption audit
‘Before rewriting this, list the three missing details that would most change your answer. If none are essential, proceed without inventing them.’
Separate structure from polish on complicated work
For complex output, asking for the polished final version too early can hide weak reasoning underneath fluent prose. First ask for the structure: decision criteria, outline, comparison categories, unresolved questions, or proposed sequence.
Once the structure looks right, ask the assistant to turn it into the finished artifact. You get a cheaper checkpoint before spending time polishing the wrong thing.
- Plan before prose.
- Comparison schema before recommendation.
- Decision list before executive summary.
- Source extraction before interpretation.
Create a repair vocabulary you can reuse
People often restart because they do not know how to describe what is wrong. A small repair vocabulary makes iteration faster and more consistent.
- Too generic → ask what source detail it ignored.
- Too confident → require facts, assumptions, and unknowns to be separated.
- Too verbose → name the decisions the output must help you make.
- Wrong tone → provide one short example of your own writing or concrete do/don't constraints.
- Over-planned → ask for the smallest useful next step and a stopping point.
Repair instead of restart
‘Keep the same facts and structure. The problem is that you turned a preference into a commitment. Restore the uncertainty and show me exactly what changed.’