Prompt engineering in 30 minutes: the foundation that lasts

Prompt engineering became a sellable topic and therefore full of jargon. The truth is simpler: clear technical writing with five patterns solves 80% of cases. The other 20% is context engineering, harness engineering, fine-tuning — separate topics.

This is what you need in 30 minutes, no zero-shot vs few-shot vs chain-of-thought vs tree-of-thought — just what works in Claude, GPT, Gemini, Llama since 2024 and still holds in 2026.

The 5 fundamental patterns

1 · Say what you want, not what to avoid

“Don’t use technical jargon” is weak. “Use language a non-technical manager would understand” is strong. Models complete patterns; positive patterns are easier to follow than negative.

Same principle in decisions: “If category is X, return ‘approved’; otherwise return ‘human review’” beats “Don’t approve if there’s doubt.”

2 · Persona handles tone; structure handles format

For tone, define persona: “You are an HR director with 15 years experience. Formal but not bureaucratic tone.” For format, define structure: “Return in 3 bullets, each one sentence, sentence starting with imperative verb.”

Persona without structure delivers OK tone in variable format. Structure without persona delivers OK format in wrong tone. Use both.

3 · Examples beat abstract instruction

“Write polite email” is vague. “Write an email following this example: [4-line example]. Same tone, same length, same formality level.” is surgical.

For recurring tasks, keep a library of 1-2 examples per context. Costs a bit in tokens, pays in consistency.

4 · Restrict the output explicitly

Models tend to be verbose. “Answer in up to 50 words” works. “Just give the direct answer, no explanation” works better. “Output JSON with schema {name, value, justification}” works even better for integration.

For programmatic use: always force schema. Without schema, output varies between calls.

5 · When wrong, show how it was wrong

Simple iteration: you say what you want, model delivers, it’s wrong, you complain. Works better if you quote the error and explain: “Your previous answer used ‘stimulating’ (line 2). That adjective isn’t in our brand guide; replace with ‘effective’.” Beats “wrong tone, redo.”

The 3 anti-patterns

Anti-pattern 1 · “Act as world’s leading expert”

Inflated persona doesn’t improve the answer. “You are the best AI expert in the world” is worth as much as “you are an AI expert.” The model isn’t more motivated — it’s just receiving a generic persona.

Use specific, contextual persona (“HR director with LGPD compliance focus”), not inflated.

Anti-pattern 2 · “Think step by step” in every prompt

Chain-of-thought had real value in 2023 with GPT-3.5 and similar. In modern models (Claude Sonnet 4.6+, GPT-5+, Gemini 2.5+), the model does this by default when the task calls for it. Forcing it hurts simple tasks — adds noise.

Use explicit CoT only in complex tasks where you want to see reasoning for debug.

Anti-pattern 3 · Single gigantic prompt with everything

Trying to solve 5 problems in the same prompt delivers 5 mediocre answers. Decompose into 5 calls (or an agent). Tokens spent, quality up.

Universal template (memorize or paste in Notion)

[CONTEXT] I'm working on <situation>. Goal is <result>.

[PERSONA] Act as <specific role, contextual flavor>.

[TASK] <What you want done>.

[INPUT] <Material/data>.

[FORMAT] <How you want the output>. Example: <short example>.

[CONSTRAINTS] <What NOT to include, length, etc>.

In 80% of corporate tasks, this template + 30 seconds of personalization delivers consistent results.

When prompt engineering isn’t enough

If you need the model to remember past conversations, search your documents, call external tools, or handle sensitive data — you’ve left prompt engineering and entered context engineering or harness engineering. Continuing to “improve the prompt” is sifting water.

Next: Context engineering: beyond the prompt — where most real 2026 gains live.