All posts

How to write better ChatGPT prompts (without memorizing a framework)

August 4, 2026 · 6 min read · By Khalid

TL;DR

  • Put the actual request in the first line, not after two sentences of setup.
  • Name the output you want — a table, five bullets, 200 words, JSON.
  • Give the model the context it can’t guess: audience, constraints, what you already tried.
  • Show one example of a good answer when the format matters.

A better ChatGPT prompt is usually the same request with three things added: what you actually want, the shape of the answer, and the context the model can’t see. You don’t need a framework with an acronym. You need to stop skipping the parts you already know matter.

Most weak prompts aren’t wrong — they’re vague. The model fills the gaps with the most average guess, and you get an average answer. Here are the four fixes that close most of that gap, in the order they matter.

1. Lead with the request, not the backstory

Put the thing you want in the first sentence. Models weight the start of your message heavily, and so do you when you re-read it. If the ask is buried under three lines of context, both of you lose the thread.

Instead of "I’ve been working on a launch email and I have a few ideas but I’m not sure about the tone and…", open with "Write a 120-word launch email for a paid Chrome extension." Then add the context underneath. The request comes first; the setup supports it.

2. Name the output you want

Tell the model the format before it picks one for you. "Give me five bullet points", "return a two-column table", "answer in under 200 words", "output valid JSON with keys title and body". Format is the single highest-leverage word you can add to a prompt, because it removes the guess that wastes the most re-runs.

This is also the fix people skip most. It feels obvious once written down, and it’s the first thing missing when a prompt returns a wall of text you didn’t want.

3. Give it the context it can’t guess

The model doesn’t know who the answer is for, what you’ve already tried, or the constraint that rules out the obvious option. Say it. "For a non-technical founder", "we’re on a free tier so no paid APIs", "I already tried X and it broke because Y". Each line removes a whole category of wrong answer.

A quick test: read your prompt as if you were a competent stranger. If a smart person couldn’t do the task from your words alone, the model can’t either.

4. Show one example when the format matters

When you want a specific style or structure, one example beats a paragraph of description. Paste a good answer and say "match this format". This is the fastest way to get consistent output for anything you’ll ask repeatedly — commit messages, product descriptions, support replies.

Weak promptWhat’s missingStronger version
Help me write a cover letterRole, job, length, toneWrite a 200-word cover letter for a junior data-analyst role at a fintech startup. Confident, not corporate.
Summarize this articleLength, format, audienceSummarize this article in five bullets for someone who hasn’t read it. One sentence each.
Fix my codeLanguage, error, what you triedThis Python function throws a KeyError on empty input. Here’s the code and the traceback. What’s wrong and how do I fix it?
The same request, weak vs. strong

Why the first prompt is usually the worst one

Under time pressure, everyone sends the half-formed version first, reads a mediocre answer, then spends three more messages fixing what they should have said upfront. The knowledge isn’t the problem. The moment is — nothing catches the gap while the prompt is still yours to change.

That’s the entire reason BeforePrompt exists: it reads your draft as you type and flags the one or two fixes that matter before you hit send, not after the bad answer comes back. But the habits above work with or without it. Start with format and context; you’ll feel the difference on the next prompt.

Frequently asked

What is the single most effective thing to add to a prompt?

The output format. Telling the model exactly what shape you want — a table, five bullets, 200 words, JSON — removes the guess that causes the most re-runs, and it’s the fix people skip most often.

Do prompt frameworks actually help?

They help you remember the parts, but you don’t need to memorize one. Most of the value comes from four habits: lead with the request, name the format, add the context the model can’t see, and show an example when the structure matters.

Why does my first prompt usually get a mediocre answer?

Because the first draft is almost always the vague one. You send it before adding format and context, get an average answer, then fix it over several follow-up messages. Adding those details up front replaces that whole back-and-forth.

Catch the gap before you send

BeforePrompt flags the missing piece while the prompt is still yours to change.