How to improve an AI prompt
Most prompts start as a rushed sentence: "summarize this," "fix this code," "write something about coffee." AI answers them, but the answer comes back generic, because the request was generic. Improving a prompt means adding back the details you already have in your head, so the model has enough to work with instead of guessing at what you meant.
Why a first draft prompt falls short
- It skips the audience. "Explain this" reads differently to a beginner than to an expert, and the model needs to be told which one you mean.
- It skips the format. Without a shape to fill, the model picks one for you: often a wall of paragraphs when you wanted three bullets.
- It skips the length. "Summarize this" could mean one sentence or one page. Say which.
- It skips the constraints. Keep this word, avoid that phrase, stay under a word count: spell these out so the model knows they matter.
Before and after: three examples
Writing. Before: "Write a blog post about coffee." After: "Write a 400 word blog post for home baristas about dialing in an espresso grinder. Use a friendly, practical tone, three short sections with headings, and end with one thing to try today."
Code. Before: "Fix this code." After: "This function should return the total price of the cart but returns NaN. Here is the function and the error message. Explain what is wrong, then show the corrected code."
Summarizing. Before: "Summarize this." After: "Summarize this for a manager who has not read it: three sentences on what happened, what it means, and what they need to decide."
Research. Before: "Look into our competitors." After: "List five direct competitors to a $12 a month habit tracker app. For each, give the price, the one feature they lead on, and a weakness their reviews mention. Table format."
Each "after" adds the same four things: who it is for, what shape the answer should take, how long it should be, and any rule the answer has to follow. Nothing about the underlying idea changed. The prompt just stopped making the model guess.
A simple way to improve any prompt
Read your prompt back and check it against a short list:
- Who is this for? A teammate, a customer, a beginner, an expert.
- What shape should the answer take? Bullets, a table, a short paragraph, code with comments.
- How long should it be? A number, a sentence count, or "as short as possible."
- What must it include or avoid? A tone, a fact to keep, a word to drop.
- What does the model need to see? The error message, the original text, the part you are unsure about.
Answer even two or three of those and the prompt already improves. Answer all five and you get an instruction worth saving and reusing, instead of retyping a rough version every time.
Improve it once, keep it improved
A one time fix helps a single answer. Saving the improved wording helps every answer after it. Long AI chats make this matter more, not less: as the context window fills, the model leans harder on whatever you just sent, so sending a sharp, complete prompt again and again keeps the quality steady instead of drifting as the conversation goes on. That is the difference between improving a prompt once and improving your prompts, plural, for good.
How Prompt Petal's Improve does it
Improve takes a petal's instruction and rewrites it for you, following the same idea: fill in what a vague request left out. In the free version, it asks the AI already built into your device: Apple Intelligence on iPhone, iPad and Mac, or Gemini Nano on supported Android phones. On any other device, Improve copies a ready request for you to paste into Siri, Gemini, Copilot or ChatGPT instead. Either way, you get back a rewritten prompt and three choices: Replace the petal with it, Save it as a new petal alongside the original, or Try again for a different rewrite.
Pro removes the device limit: add your own ChatGPT, Claude or Gemini key and Improve calls it directly. Claude Desktop, Cursor and VS Code can also improve and save your prompts themselves, through MCP, without opening Prompt Petal at all.