Prompt Engineering: The Techniques I Actually Use Daily

In short: Strong prompt engineering rests on a few practical techniques that visibly improve outputs across ChatGPT, Claude and Gemini. A good prompt combines five blocks — a role, context, a specific task, an output format, and constraints. Adding chain of thought and few-shot examples further sharpens the model's responses.

I've been writing prompts seriously since late 2022, and I've been training teams in prompt engineering since 2023. This article condenses what I actually use day to day — not an exhaustive list of academic techniques, but the ones that visibly change the quality of my outputs on ChatGPT, Claude, Gemini and the open-source models. If you want to gain 50% in output quality right away, these five techniques are enough.

What Makes a Good Prompt, in Five Blocks

My typical prompt contains five elements, and I put them in this order. A role: "You are a GDPR expert." The model adapts its register and its precision. A context: the situation, the constraints, what already exists. A specific task: a clear action verb (write, analyze, compare, list, summarize). An output format: "in 3 paragraphs", "as a table", "in JSON". Constraints: length, tone, technical level, language.

When a participant in a training session shows me a prompt that didn't work, nine times out of ten it's missing two or three of these blocks. The density of information in the prompt is what makes the difference.

Technique 1: Chain of Thought

I ask the model to "think step by step" before giving a final answer. It's especially effective for complex problems, calculations, and multi-step reasoning. On Claude, I've measured a 20-30% gain in precision on legal analyses simply by adding "explain your reasoning step by step before giving the answer."

Anthropic documents this technique with detailed examples, and OpenAI does the same for GPT. Recent models partially integrate this reasoning natively, but making it explicit still helps.

Technique 2: Few-Shot (Learning from Examples)

I give two or three examples of the expected output before asking the real question. The model grasps the pattern and reproduces it. This is by far my most effective technique for repetitive tasks or stylistically precise ones.

A concrete example I use to rewrite SEO titles:

`` Here's how to rewrite article titles in SEO mode: Original: "How does AI work?" → SEO: "Artificial Intelligence: Complete Guide 2026" Original: "The advantages of DeepL" → SEO: "DeepL: 7 advantages that make it the best AI translator" Now rewrite: "Everything about Mistral AI" ``

Three examples are enough in most cases. Beyond that, the marginal gain becomes small and you're eating into your context for nothing.

Technique 3: Maximum Contextualization

LLMs have no context about your situation by default. The more I give, the better the response. It's the simplest technique to apply but the most neglected.

Bad prompt: "Write a follow-up email."

Good prompt: "Write a B2B follow-up email for a SaaS prospect who downloaded our white paper 7 days ago. He's a CMO at a 50-person French scale-up. Warm but professional tone, 3 paragraphs, call to action toward a 20-minute demo. No forced formality. Sign it Laurent."

The second prompt produces an email I can send almost as-is. The first produces a generic, unusable draft.

Technique 4: Decomposing Complex Tasks

For long and complex tasks, I break them into several sequential prompts rather than one massive prompt. My typical flow for an article: first "generate a 5-part outline for an article on [topic]", then "write part 1 of the following outline: [outline]", then "write part 2", and so on. The final result is higher quality than if I ask for the complete article all at once, and I keep control of the structure.

This technique is also the foundation of anything that resembles an autonomous agent or complex reasoning chains.

Technique 5: Iteration and Refinement

A first prompt rarely gives the perfect result. I iterate systematically. "Rewrite it shorter", "make it more formal", "add concrete examples", "remove the clichés", "challenge this conclusion." Iteration is where the final quality is decided.

A bad reflex I fight against in training: starting a prompt over from scratch instead of iterating. Keeping the context of the conversation is almost always more effective than starting over.

Prompts for Specific Use Cases

The patterns I've stabilized. For SurferSEO + AI: "write an SEO-optimized H2 for the query [keyword], length 5-8 words, natural, no stuffing." For Jasper AI: integrate the Brand Voice into the context of the system prompt. For GitHub Copilot: explanatory comments before a function guide the completion better than silence. The more you describe what the function should do in a comment, the better the suggestion.

What LLMs Don't Do Well Despite a Good Prompt

Knowing the limits saves you useless prompts. Calculating complex numbers precisely: I use Code Interpreter or I ask the model to produce Python that I run myself. Accessing real-time data: I go through Perplexity or tools with web access. Remembering previous conversations beyond the context window: everything has to be re-injected into the prompt.

No prompting technique gets around these limits. You have to pair the right tools with the right tasks.

My Reading for Trust-Vault

Prompt engineering is a cross-cutting skill that improves the use of every tool in my catalog. To evaluate each tool's prompt-comprehension capabilities, I systematically test with a standardized protocol before scoring on the Reliability pillar of my Trust Score methodology.

My closing advice: mastering these five techniques already puts a user in the 10% who really get the most out of their tool. The rest comes with practice and from watching what works on your specific cases.

--- Sources: Anthropic — "Let Claude Think" prompt engineering guide; OpenAI prompt engineering guide; Wei et al. — "Chain-of-Thought Prompting Elicits Reasoning", 2022; Brown et al. — "Language Models are Few-Shot Learners", 2020.