How Do You Tell Claude the Result You Want?
Describe the finished thing, who will use it, and how you'll judge it. Claude can work toward a result it can see, so "done" should be something you could actually check, not a topic or a verb.
I need [the finished thing] for [who will use it], so they can [what they'll do with it]. It's done when: - [a check you could apply, such as "under 200 words"] - [a second check, such as "names the decision needed by Friday"] - [what's out of scope] Background: [only the facts that change the answer] Before you finish, compare your result to the list above and tell me about any item you didn't meet.
- Topica rejection email
- Taskwrite a rejection email
- Resultshe applies again next year
- Result you can checktwo short paragraphs, the real reason, no empty promises
Try it now
- 1Take a task you're about to give Claude, and write one sentence that starts "It's done when". If you can't finish it, work that out first.
- 2Paste the prompt above with two or three checks, and send it.
- 3Read the end of the reply. You'll know it worked when Claude reports against your list, and you can see at a glance which items it met.
Use it when
Whenever the output has a job to do: a document someone will act on, a function that has to pass a test, or an answer that feeds a decision. It helps most when you've been getting answers that are correct but not useful.
Skip it when
When you're exploring. Anthropic's Claude Code guide says a loose prompt like "what would you improve in this file?" is fine, as long as you accept making more corrections along the way. Simple factual questions don't need it either.
Why it works
Anthropic's Claude Code guide says Claude tends to read a request conservatively and build the smallest thing that satisfies it, because it can't see how much you were hoping for. The general prompting docs make the same point: if you want more than the task strictly requires, ask for it. Their example is "create an analytics dashboard," which improves once it asks for a fully featured result. For research, the docs recommend defining what a successful answer looks like. Wording such as "think step by step" won't do this job. The Claude Code guide calls those phrases ordinary prompt text and says the effort level is the main control for how hard Claude thinks. Your words are better spent on the result.
Example
"Write a rejection email" tells Claude the format. Compare this: "I need an email to a finalist we didn't hire, so she'll apply again next year. It's done when it's two short paragraphs, gives the real reason, which is that we chose someone with more infrastructure experience, and promises nothing we can't deliver." The second version gives Claude what to aim for and what to leave out.
Going further
In Claude Code, you can turn your definition of done into a check. /goal all tests in tests/auth pass and lint is clean sets a condition that a separate model checks after every turn, and Claude keeps going until it's met. The docs say this matters mostly for long background runs. To widen scope, the guide says one phrase still helps: "be ambitious," followed by what you want, such as empty states and keyboard navigation. Also note the order for long inputs. Anthropic's docs put long documents at the top of a prompt and the query below them, so with a big paste your result statement should come after it.
Common mistakes
- Confusing the task with the result. "Write an email" is a task, and "get a yes or no by Friday" is a result.
- Listing checks no one can apply, such as "make it engaging." Replace them with something you could verify, such as one concrete example per section.
- Expecting extras you didn't mention, then blaming the model for a minimal result. Anthropic's docs say to ask for them.
- Adding "think hard" and leaving the result vague. In Claude Code, use
/effort highwhen the task needs more thought, and state the result in your prompt.
Sources
Checked 6 November 2026. Products change often, so confirm details with the vendor.