ChatGPTPrompting mistakesLevel 1: Basic prompts

Why Does ChatGPT Ignore Parts of a Long Prompt?

Usually you've given it too many jobs and rules at once, so some get less attention. Length isn't the problem. A long source document is fine, but ten separate tasks in a message aren't. Split them into steps, and keep only the one or two rules that matter most. This is a rule of thumb, based on OpenAI's advice below.

Prompt to copy
Don't do this task yet. Here's everything I want:

[paste your long prompt or messy list of requests]

Split it into the fewest steps that make sense, in order. For each step, give me:
1. A one-sentence goal
2. What it needs from me or from the previous step
3. How I'd check that it's right

Mark any of my rules that conflict, and ask me which one wins. Also tell me which matters more for this work: [accuracy / creativity / speed].

Then stop. When I reply "go", do step 1 only and wait for my OK before the next one.
  1. 1Comparison table with sources
  2. 2You check it
  3. 3200-word email from the table
  4. 4Then the translation
This card's example as a short chain. A rule of thumb. OpenAI's Academy advises breaking multi-part requests into smaller steps.

Try it now

  1. 1Find a long prompt that disappointed you, and paste it into the prompt above in New chat.
  2. 2Read the steps and settle any conflicts it flags, then reply "go".
  3. 3Check step 1 before you approve it. It worked if every rule from your original prompt ends up in a step, and you can correct one step without redoing the rest.

Use it when

When your prompt has several verbs ("research, outline, draft, translate and format"), when a rule you set keeps getting dropped, or when a single answer needs several different kinds of checking.

Skip it when

Don't split a single task with a lot of supporting detail. A long prompt that gives one goal plus the facts it needs is fine. And when the parts depend on each other, like "the summary must match the table", keep them together, and ask for a check at the end.

Why it works

OpenAI's Academy says to break multi-part requests into smaller steps, and to set priorities by telling the model whether accuracy, creativity or speed matters most. Its Prompting guide says to "use only the parts that help," and to focus on the one or two boundaries that matter most.

Splitting also helps you. With one task per message, you can see which step went wrong and correct just that. A later step can then build on work you've already approved, so a mistake in the outline doesn't get copied into the draft and the summary.

Some big jobs are really a sign you're in the wrong mode. OpenAI says Chat suits quick questions and drafts, while ChatGPT Work suits tasks that are multi-source, multi-step or produce files. Free and Go have limited Work in the desktop app, and Plus has it on desktop, web and mobile.

Example

You send: "Research our three competitors, write a comparison, draft a sales email, translate it into German, and keep it under 200 words in a friendly tone." The 200-word limit could apply to any part, and the research can't be checked if it's buried in an email. Split it: first a comparison table with sources, which you check. Then a 200-word email based on the approved table. Then the translation.

Going further

For a repeatable chain, keep the steps in one chat, or use a project when several chats should share files and instructions. On Go and up, OpenAI suggests refining a recurring prompt in a normal chat first and then scheduling it inside that chat. For a truly long task in Work or Codex, write one brief and let the agent do the stepping. OpenAI's Long-running work page says /goal starts Goal mode, with the goal text acting as both the first prompt and the completion criteria, and to use /plan first if the outcome is unclear. Keep in mind that OpenAI warns agents can take longer and cost more, so check your usage first.

Common mistakes

  • Putting ten rules in one line, then not noticing which ones were dropped. Number them, and ask for a check at the end.
  • Splitting tasks that depend on each other, so the summary no longer matches the table.
  • Adding a rule for every past disappointment. If it applies across chats, it belongs in custom instructions, not in each prompt.
  • Moving to the next step without checking the last one, so a wrong fact gets copied everywhere.

Sources

Checked 6 November 2026. Products change often, so confirm details with the vendor.

Related

Beyond the course