How to Write Better AI Prompts: 9 Techniques That Work

Prompting is the 80% skill: technique matters more than which assistant you use. These 9 techniques work across ChatGPT, Claude, Gemini and every other chatbot — each shown with a weak prompt, a strong prompt, and why the difference matters. Practice on any free tier (our alternatives roundup lists them).

Nine techniques, each with a before-and-after example. Photos: Pexels (free license).

Quick answer: assign a role, supply context it cannot guess, specify format, show an example, set constraints, iterate in conversation, make it question you first, verify numbers — and save what works as reusable instructions.

Contents

• 1–3: Role, context, format · 4–6: Examples, constraints, iteration · 7–9: Questions-first, verification, saving
• Full before-and-after · Mistakes · FAQ

Techniques 1–3: role, context, format

1. Assign a role. Weak: "Help me negotiate." Strong: "You are a procurement negotiator with 15 years in manufacturing…" Roles focus vocabulary, standards and stance.

2. Supply unguessable context. Audience, constraints, background, what you already tried. The model cannot know your manager hates bullet points — tell it once per conversation.

3. Specify format. Length, structure, tone, ending: "120 words, 3 bullets, direct tone, end with one question." Format instructions are the cheapest quality lever that exists.

Techniques 4–6: examples, constraints, iteration

4. Show, don't just tell. Paste 2–3 examples of good output ("match this style/tone/structure"). Few-shot examples beat paragraphs of description.

5. Set constraints. Budget, words, reading level, forbidden phrases ("no hype words, no exclamation marks, SI units only"). Constraints force decisions the model would otherwise dodge.

6. Iterate conversationally. "Shorter." "Simpler words." "Give 3 variants of paragraph 2." Five 10-second refinements beat one perfect first prompt — conversation is the interface, not a fallback.

The whole skill in one sentence: say what good looks like. Photos: Pexels (free license).

Techniques 7–9: questions-first, verification, saving

7. Make it question you first. "Ask me 5 questions before drafting." Transfers the burden of completeness to the party with infinite patience — dramatically better first drafts for complex tasks.

8. Demand verifiability. "Cite sources for every statistic" / "flag anything you're unsure about." Doesn't eliminate hallucination; makes it checkable. Then actually check the three claims that matter.

9. Save what works. Working prompt → custom instructions, Gems, custom GPTs, or a notes file. Prompt libraries compound: see setup guides for ChatGPT and Gemini.

Full before-and-after

Task: investor update email. Before: "Write an investor update." → generic template, wrong metrics, wrong tone. After: "You are a startup COO writing to seed investors. Context: SaaS, $42k MRR (+8% MoM), churn 3.1%, runway 19 months, hiring 2 engineers. Format: 150 words, wins/challenges/ask sections, confident but honest tone, end with next update date." → usable draft needing only personal touches. Same model, same task — the difference is entirely prompt.

5 mistakes that waste the techniques

1. One-shot expectations: judging by the first answer instead of the fifth.
2. Context starvation: expert model, intern-level briefing.
3. Trusting numbers: dates, prices, citations — verify, always.
4. Remaking setup: retyping preferences instead of saving instructions.
5. Blaming the model: switching assistants monthly instead of fixing prompts once (see Claude vs Gemini — the gap is smaller than your prompting gap).

3 practice drills (15 minutes total)

Drill 1 — rewrite: take your last vague prompt, add role + context + format, compare outputs side by side. Save the winner's pattern.

Drill 2 — constrain: give the same task with three different constraint sets (100 words formal / bullets casual / table analytical). Notice how constraints, not adjectives, control quality.

Drill 3 — questions-first: on your next complex task, open with "ask me 5 questions first" and answer honestly. Compare the draft against your usual one-shot results.

Do all three once and prompting stops being theory — it becomes reflex.

Frequently asked questions

Do I need to learn prompt engineering?
Not as a discipline — these 9 habits cover ~95% of real gains. Courses help power users; everyone else needs practice, not theory.

Do the same prompts work on all AI tools?
Largely yes. Role/context/format/examples transfer across ChatGPT, Claude, Gemini and the rest; only app-specific features (Gems, custom GPTs) differ.

How long should a prompt be?
As long as the unguessable context requires — usually 3–6 sentences. Longer than a paragraph often means the task should be split or exemplified instead.

Can AI detect a good vs bad prompt?
Technique 7 exists for this: ask the assistant to critique or improve your prompt before answering. Iterating the question beats answering the wrong one brilliantly.

Where do I save reusable prompts?
Custom instructions/memory (ChatGPT), Gems (Gemini), Projects (Claude), or a note in your system — organized in tools from our Notion vs Obsidian guide.

Keep exploring: How to Use ChatGPT · How to Use Gemini · Claude vs ChatGPT · About this site

Models change; technique compounds — these habits survive every model generation.