Guide ⚙️ Prompt engineering part 2: 7 techniques that separate pros from passengers

Part 1 covered the basics. This is the next level — the seven moves I see pro prompters use that almost nobody talks about. Each with a mini-example you can steal on the spot.

1. Stack Role + Context + Task (in that order)
Amateurs: "write a product description".
Pros: role, then context, then task.

Code:

You are a D2C email copywriter for premium Indian skincare brands.
Context: festival sale, 25% off, stock limited to 200 units, playful tone.
Task: write a 90-word email body with one CTA.

The order is not picky — it forces you to give all three. Missing context is why outputs feel generic.

2. Show, don't describe (few-shot)
One example beats a paragraph of adjectives. Give the model YOUR pattern:

Code:

Match this style exactly:
Input: "moved to cloud"
Headline: "Company cuts server costs 40% with cloud migration"
Now do: "hired 50 engineers"

3. Hard constraints, not vibes
"Keep it short" is a vibe. Constraints you can count are hard:

Code:

Rules: exactly 3 bullets. Each under 10 words. No adjectives. End with a question readers must answer yes/no.

Then — key habit — READ the output and check it. Models break counting rules more than any other.

4. Chain-of-thought is a tool, not a religion
"Think step by step" helps math, logic and debugging. It HURTS creative writing and summarising (makes them bloated). Match the trick to the task: reasoning → yes, step-by-step; creativity → ask for 3 drafts instead.

5. Force structured output
Don't hope the answer is parseable — demand a shape:

Code:

Return ONLY a table: Column A = task, B = owner, C = deadline, D = risk (one word). No intro, no outro, no markdown fences.

This is the technique that makes prompts usable inside automations — the 30-day roadmap leans on it heavily.

6. The self-critique loop (the secret one)
Two-step prompt that upgrades everything:

Code:

Draft the answer. Then, as a harsh critic, list 3 specific weaknesses in your own draft. Then produce the final version fixing all three.

Costs one extra call. Saves you a full round of prompting. This is 80% of "why does his AI write better than mine".

7. Version your prompts like code
Save prompts in a note with names + dates (my_format_v1, my_format_v2). When a model updates and your prompts suddenly misbehave, you can roll back instead of rewriting from memory. Pros keep a library; passengers re-type from scratch.

Where to practice these tonight
Pick technique #2 or #6 — they're the highest-payoff — try them on a real task, then post your before/after in Prompt sharing. I read every experiment. And if you missed the starter pack, grab the 25 copy-paste prompts first. ⚡

Critic-then-rewrite is my default. Bonus trick: have it list the 3-5 core assumptions it's making before answering. Saves me from garbage outputs on unfamiliar topics by catching bad premises early.