Prompt Engineering

Prompt engineering tips & techniques that actually move the needle.

The patterns that lift output quality across every LLM — explained in plain English with copy-paste examples. No fluff, no “ultimate” lists.

Try the free tool
Add to ChromeAdd to EdgeAdd to Firefox
4.9 / 5 average ratingFree foreverNo account requiredWorks in ChatGPT, Claude, Gemini & more
Ranked by ROI

Ten prompt engineering tips, ranked by ROI.

The first five apply to every prompt you write; the last five are situational levers worth knowing for hard tasks.

  1. 01Name the audience explicitly

    "Explain X to a senior engineer" lands very differently from "explain X". Audience changes vocabulary, depth and assumed context in one move — highest-ROI single edit you can make.

  2. 02Specify exact format

    "In 5 bullets", "in a 3-column markdown table", "in exactly 120 words". Format is the highest-leverage constraint you can add. Cuts variance ~70%.

  3. 03Use few-shot examples

    One or two examples of the style or output you want. Doubles quality on creative, classification and code tasks — beats any amount of descriptive language.

  4. 04Add chain-of-thought for reasoning

    Append "Think step by step before answering" for any task involving math, logic or multi-step planning. Reliable +15-25% accuracy on benchmarks.

  5. 05Use role prompting

    "You are a senior backend reviewer..." primes vocabulary and depth without changing the model. Cheapest tone control there is.

  6. 06Add explicit negative constraints

    "No marketing language", "no disclaimers", "no fabricated citations". Prevents the most common failure modes upfront — and they’re what users complain about most.

  7. 07Separate context from task

    Use clear delimiters (---, ```, XML tags) between the source material and the instruction. Claude particularly rewards XML.

  8. 08Ask for structured output

    "Return as JSON with keys X, Y, Z" produces parseable, predictable output every time. Essential when chaining LLMs in code.

  9. 09Use self-consistency for hard tasks

    Prompt the model to generate 3 answers and pick the best one. Trades cost for accuracy on multi-step reasoning, math, evaluation.

  10. 10Iterate with diffs, not rewrites

    When refining output, give the model the current version and tell it what to change. Faster and more accurate than re-prompting from scratch.

Top of mind

Six tips that never get old.

Audience over adjectives

Saying "for a CFO" is worth more than ten adjectives like "concise, professional, executive-friendly".

Length caps tighten everything

"In 120 words" or "in 5 bullets" reduces output variance by ~70%. The model stops padding.

One example > ten adjectives

A single example of the style you want changes the output more reliably than any descriptive language.

Constraints prevent failure modes

Explicit negatives ("no fluff", "no fabricated facts") catch 80% of common errors before they happen.

Chain-of-thought for hard tasks

"Think step by step" reliably boosts reasoning quality on math, logic and multi-step problems.

Structured output for pipelines

When chaining LLMs in code, always demand JSON or XML. Free-form output breaks pipelines.

Prompt Engineering Tips FAQ

Common questions.

For everyday use, specifying audience and exact output format wins by a wide margin — they explain most of the variance in answer quality. For reasoning-heavy tasks, chain-of-thought ("think step by step") is the highest single lever.
Zero-shot for clear, well-scoped tasks. Few-shot when the style, format or category structure is hard to describe in words but easy to show with 1-2 examples.
Yes, it works on GPT-4o, GPT-5, Claude 3.5 / Claude 4, Gemini 2.0 and most open models. Newer reasoning models (o1, o3) do CoT internally, so explicit "think step by step" matters less for them.
For hard reasoning tasks, generate 3-5 answers and pick the most common one. Trades cost for accuracy. Implement either by prompting the model to "give 3 attempts" or by sampling N times via API.
Yes. AI Prompt Fixer applies the most reliable patterns — audience, format, constraints, examples — automatically, with a one-click rewrite in ChatGPT, Claude, Gemini and Perplexity.

Apply these prompt engineering tips automatically.

AI Prompt Fixer encodes the patterns above as a real-time scorer + rewriter. Paste any prompt and see them applied in one click.

Try the live demo
Add to ChromeAdd to EdgeAdd to Firefox