How to Write Better AI Prompts: 9 Tactics That Work

TL;DR: Give context, set constraints, show an example, demand a structure, and iterate in small steps — these five habits plus four advanced tactics below improve output on every chat model.

Prompt quality is the cheapest performance upgrade in AI. The same model produces vague filler from a three-word request and genuinely useful work from a well-built prompt. The nine tactics below are ordered from highest leverage to most specialized, each with an example you can copy and adapt.

Foundations: context, constraints, and roles#

The first three tactics fix the large majority of bad outputs. They all answer the question the model cannot ask: what situation is this for, what limits apply, and whose perspective should shape the answer.

Tactic 1: supply context, not just the task#

Bad prompts describe the task ("write a launch email"). Good prompts describe the situation: audience, product, stage, and what success looks like. Context lets the model choose appropriate specifics instead of generic ones.

Weak: "Write a launch email for my app."
 
Strong: "Write a launch email for Taskbeam, a $12/month task app for
freelance designers. Audience: 2,000 newsletter subscribers, mostly
trial users. Goal: 5% click-through to the new timeline feature.
Tone: friendly, concrete, under 150 words."

Tactic 2: set explicit constraints#

Constraints are quality controls: length caps, reading level, forbidden words, required sections, and format. Models follow stated limits well, but only the ones you state — every unspecified dimension gets filled with the most average choice available.

Tactic 3: assign a role with standards#

"Act as a senior technical editor who rejects vague claims" outperforms "act as an expert" because it names the standard, not just the title. Pair the role with its checklist: what would this professional check, and what would they refuse to accept.

Takeaway: Context picks the specifics, constraints bound the output, and a role with standards sets the bar — these three fix most weak prompts.

Structure: examples, formats, and step-by-step work#

Once foundations hold, structure tactics raise consistency. They control the shape of the answer so outputs arrive ready to use instead of ready to reformat.

Tactic 4: show one example of good output#

A single input-output example teaches format faster than paragraphs of description. Show the model a headline you like, a summary at the right density, or a table with the columns you want — then ask for the same treatment of new material.

Tactic 5: demand a structure#

Name the sections, the order, and the length of each. "Respond with: verdict (1 line), three reasons (1 line each), then risks (bullets)" beats hoping the answer lands in a usable shape. Structures also make follow-up edits surgical: "redo reason 2 with a concrete number."

Tactic 6: split complex jobs into steps#

Ask for the outline first, approve it, then draft section by section. Step-by-step prompting keeps long outputs coherent because each step's result becomes the next step's verified input — the same reason writers outline before drafting.

Takeaway: Examples teach shape, structures make output usable, and steps keep long work coherent — never ask for outline, draft, and polish in one breath.

Refinement: critique loops and iteration discipline#

The final three tactics govern what happens after the first answer. Most people rewrite the whole prompt when one paragraph disappoints; disciplined iteration keeps what works.

Tactic 7: ask the model to critique itself#

"List the three weakest claims in your answer and what would verify each" catches invented specifics surprisingly well. Follow with "rewrite section 2 replacing unverified claims with flagged placeholders." Verification becomes part of the workflow instead of an afterthought.

Tactic 8: change one thing per retry#

When output is close, adjust a single variable — tone, length, one section — and keep everything else fixed. Wholesale rewrites discard the paragraphs that were already right and randomize the rest. Small steps converge; big jumps wander.

Tactic 9: save winners as reusable templates#

Prompts that work are assets. Store the exact wording (small changes shift outputs), note which model and settings it ran on, and build a small library per repeating task: summaries, outreach, code reviews, lesson plans. Compounding returns come from reuse, not from writing fresh prompts daily.

• Keep prompts under 200 words; split bigger jobs into steps.
• State the most important constraint first and repeat it last.
• Save exact winning wording with model name and date.
• Ask for flagged uncertainty instead of confident guesses on facts.

Takeaway: Critique loops catch errors, single-variable iteration converges, and saved templates turn one good prompt into a permanent capability.

Conclusion#

Better prompting is five habits — context, constraints, roles, examples, structure — plus disciplined iteration and template reuse. Practice them on this week's real tasks and every model you touch gets more useful. For background on how this site writes and verifies its guides, see the about page.

Sources#

• Prompt Engineering Guide
• OpenAI platform documentation
• Anthropic documentation