Prompt Engineering Best Practices (and the Mistakes That Undo Them)

Prompt engineering best practices aren't tricks or magic phrases — they're a short list of habits that consistently close the gap between a mediocre AI answer and a genuinely useful one. Six practices do most of the work. Six mistakes quietly undo them. This is the practical version: what to do, and specifically what not to, with the reasoning behind each so you can apply it to prompts these examples don't cover.

The six practices that actually matter

1. Assign a specific role, not a generic one

“You are a helpful assistant” does almost nothing — it's already the model's default posture. “You are a senior tax attorney specializing in pass-through entities” does a lot: it sets vocabulary, depth, and the kind of caveats a real expert would include. Specificity is the whole value of a role; a vague one is wasted words.

2. Front-load the instruction

Models weight the beginning and end of a prompt more heavily than the middle, a pattern often called “lost in the middle.” If your most important constraint is buried in paragraph three of context, it's more likely to get dropped. State the core task early, put supporting context after it.

3. Say what to exclude, not just what to include

“Write a product description” leaves the model free to default to whatever's generic for the category — hype adjectives, exclamation points, the same three adjectives every AI writes. “No hype language, no exclamation points, no superlatives” closes off the lazy option and forces something more specific to survive.

4. Specify the output format, every time

This is the single highest-leverage line in most prompts and the most commonly skipped. Table, numbered list, JSON, three short paragraphs, a specific word count — say it. Without a format instruction the model defaults to prose, which is frequently the wrong container for the information you actually need.

5. Show one example when tone or format is unusual

For anything with a standard shape (a summary, a straightforward email), skip this — it's overhead. For anything with a specific, non-obvious tone or structure, one well-chosen example outperforms a long paragraph describing the pattern. Three or four examples rarely beats one good one.

6. Iterate with targeted critique, not a full rewrite

When an output is close but not quite right, tell the model exactly what to change — “shorter, cut the second paragraph, add a specific number in the close” — instead of restarting with a longer prompt. Targeted critique keeps what was already working and fixes only what wasn't.

The mistakes that undo all six

Following the six practices above only works if you're not simultaneously making one of these. In practice, most weak prompts have one of these mistakes doing the damage, not a missing practice.

Topic instead of task

The most common mistake by far. “Remote work” is a topic. “List four specific challenges of managing async teams across time zones, with one tactic for each” is a task. A role and a format won't save a prompt that never states what kind of output it wants.

Contradictory instructions

“Be concise” followed later by “cover every angle in detail” cancels itself out — the model has to guess which one you meant more, and it usually splits the difference badly. Read your own prompt back before sending it and check for instructions that pull in opposite directions.

Negative framing instead of positive

“Don't be vague” is weaker than “write like a senior engineer explaining to a junior colleague,” because the model has to infer the positive version of your instruction from its negation. State what you want directly rather than what you don't.

Cargo-culting reasoning prompts

Adding “think step by step” to a simple task doesn't help — modern reasoning models already do this internally on anything that benefits from it, and pasting the phrase everywhere just wastes tokens without wasting quality. Reserve it for problems that genuinely need visible, checkable reasoning steps.

No context on the actual constraint

“What's the best marketing channel?” produces an answer that's technically correct for the average business, which is exactly why it's useless for yours. The model has no idea about your budget, audience, or what you've already tried unless you say so.

Restarting instead of iterating

When an output is 80% right, writing an entirely new prompt from scratch throws away the 80% that worked along with the 20% that didn't. Targeted follow-up almost always gets there faster.

Applying this without rewriting every prompt by hand

These six practices and six mistakes are the same pattern, stated two ways: role, task (not topic), constraints (including exclusions), format, and one example when needed. It's a checklist, which means it's mechanical enough to automate — try a real prompt below and see the practices applied for you:

TL;DR

• Six practices: specific role, front-loaded instruction, explicit exclusions, defined output format, one example when needed, targeted iteration.
• Six mistakes that undo them: topic instead of task, contradictory instructions, negative framing, cargo-culted reasoning prompts, missing context, restarting instead of iterating.
• Topic-instead-of-task is the single most common mistake and the highest-leverage one to fix first.
• The practices are identical across ChatGPT, Claude, and Gemini — only the emphasis shifts per model.
• Since it's a checklist, it's mechanical enough for a one-click tool to apply automatically.