Most advice on how to write effective prompts is a list of techniques. This post is organised the other way round, by the problem. There are five problems behind nearly every bad output, each has a fix, and the fixes are the same across ChatGPT, Claude and Gemini with a few differences worth knowing. If you learn to recognise the problem from the output, you will know which fix to reach for without consulting a list.
Problem 1: the prompt is vague
The output is on topic and useless. You read it and think "yes, but that is not what I needed." The prompt named a topic rather than a result, or used a verb the model can satisfy in many ways ("improve," "help," "look at").
The fix: state the result. What will exist when the model is done, for whom, how long, in what shape. "Improve this" becomes "shorten this to 120 words and remove every sentence that repeats an earlier one." A prompt with a stated result is half written.
Problem 2: the prompt is overloaded
The output does most of what you asked and drops one instruction, and which one it drops changes between runs. The prompt asked for several things at once or ran past about 300 words of instruction.
The fix: one task per prompt, and cut the instructions that have never been associated with a failure. For a task with several parts, ask for the plan first, then the parts in order. Long material is fine; long instruction is what fails.
Problem 3: the output is inconsistent
The same prompt gives a good answer on Monday and a mediocre one on Tuesday, or works on three inputs and fails on the fourth. The prompt leaves decisions to the model, and the model decides differently each time.
The fix: replace every adjective with a rule and give one example of the output's shape in its own block. "Professional tone" becomes "no exclamation marks, no first person plural, past tense." Where a setting is available, lower the temperature for tasks with one right answer. Inconsistency is almost always underspecification.
Problem 4: the context is missing
The output is correct in general and wrong for you. It describes a company like yours rather than your company, or a customer like yours rather than the one in the email. The model filled the gaps with averages because nothing in the prompt told it otherwise.
The fix: give it the facts it cannot guess, in a labelled block after the instruction: who the output is for, what has happened, what the constraints are, the material itself. And add a rule for the gaps that remain: say what is missing and stop, or state the assumption and mark it.
Problem 5: the change was never tested
You fixed the prompt for the input that failed and it now fails on a different one. Or you changed three things at once and do not know which helped. Or a model update changed the behaviour and nobody noticed for a month.
The fix: a small test set of varied inputs, a written description of a good output, one change per version, a log. This is the habit that makes the other four fixes stick, and it takes an hour to set up with a spreadsheet. The prompt optimization guide covers the loop.
Where the models differ
ChatGPT follows a conversational prompt well and benefits most from an explicit format and an explicit stop; it adds a summary and a closing offer by habit, so say when you do not want them. Newer versions take instructions literally, so what you used to leave implied now has to be written.
Claude responds best to structure: role, task, material in a labelled block, rules as a short list, format last. It follows the last instruction most closely and will ask a clarifying question when told it may, which saves a wrong first draft on long tasks. Give it the length in words.
Gemini wants the instruction at the top and the material below with a clear separator, and it follows a system instruction over the user turn strictly. For structured output, give it a schema rather than an example. For long material, put the question at the end as well as the start.
The differences are about format compliance and refusal behaviour more than quality. Run the same prompt in two models on the same inputs and the differences tell you which rules each needs spelled out.
A prompt that solves all five
Task: from the meeting transcript below, write the list of decisions and the list of actions. Context: a product team's weekly planning; the readers were in the meeting and want a record, not a summary. Material: [transcript]. Rules: a decision is something the group agreed to do or not do; quote the line it came from; an action has an owner named in the transcript or is marked unassigned; do not invent an owner. If the transcript has no clear decisions, say so. Format: two headed lists, decisions then actions, nothing else.
A result, one task, rules instead of adjectives, context and material separated, a gap rule, the format last. Test it on three transcripts before trusting it on the fourth.
If you would rather start from a working prompt than a blank one, every prompt in the prompt library is built to this shape for a specific task, and the three elements of an effective prompt is the short version of the structure behind it.
To learn by editing working prompts, open the education prompts.
