ClaudeGetting startedLevel 1: Basic prompts

How Do You Check Claude Understood Before It Starts?

Ask Claude to restate the task and list its assumptions, then tell it to stop until you reply. In Claude Code, use Plan mode instead: Claude researches and proposes a plan without editing your source, and in most setups edits stay blocked until you approve.

Prompt to copy
Here is the task: [your request].

Don't start it yet. First reply with:
1. The task in your own words, in two sentences.
2. Who you think the result is for, and how long it should be.
3. The assumptions you're making, including what you're treating as out of scope.
4. Anything you'd have to guess at, as questions I can answer in a word or two.

Then stop and wait for me to say go.
  1. 1Ask: restate, then stopinclude your assumptions
  2. 2Claude restates and waits
  3. 3You correct one lineread the out-of-scope items first
Repeat until it's right, then say go
Rule of thumb, not an Anthropic chart: a wrong assumption costs one line to fix here, before anything is written.

Try it now

  1. 1In a chat, paste the prompt above with a real task, and read the assumptions list first. That's where most misreadings show up.
  2. 2Correct anything that's wrong in a line or two, then reply go.
  3. 3In Claude Code, press Shift+Tab until the status bar shows ⏸ plan mode on, or start one prompt with /plan.
  4. 4When the plan appears, choose No, keep planning to change it, or Yes, manually approve edits to review each edit. You'll know it worked when no file changes until you approve.

Use it when

When a request could reasonably mean two things and the wrong one is expensive: a board summary you'll send tomorrow, a contract review with a specific focus, or a code change across many files.

Skip it when

Short tasks with one obvious reading, such as translating a sentence. Also skip it when you don't know what you want yet. A restatement can only reflect what you said, so ask Claude to interview you instead, as described below.

Why it works

Claude doesn't always pause by itself. Anthropic's Sonnet 5.5 guide says that on an open-ended request the model can start building a presentation, report or video when you only wanted ideas, and its fix is to say so in the request. Asking first also saves usage. Anthropic's help page says conversation length affects how quickly you reach your limits, and a long wrong answer stays in the history for every later turn. A restatement costs a few lines, and correcting it costs one.

Example

Say you ask for "a summary of this report for the board." Claude replies that it plans one page for non-technical directors, focused on financial results, and is treating the security findings as out of scope. The security findings were the reason you opened the report. You write one line to change that, then say go. You find the problem before anything has been written.

Going further

If you can't say what you want yet, flip it. Anthropic's Claude Code guide suggests asking Claude to interview you until it could write the spec, and to save the result as a file such as SPEC.md. Then start a fresh session to build from it, so work begins with a clean context and a written definition of done. In Plan mode, Ctrl+G opens the plan in your text editor so you can change it directly. One trap: the docs say that in interactive terminal sessions with bypass permissions available, Claude is still told to plan without editing, but edits it attempts can run without a prompt.

Common mistakes

  • Leaving out the stop line. Without it, Claude may restate the task and carry straight on.
  • Asking it to write out its internal reasoning instead of its understanding of the task. Anthropic's Sonnet 5.5 guide warns that such prompts invite a reasoning_extraction refusal.
  • Approving a plan after reading only the first lines. The out-of-scope items and assumptions are what you most need to read.
  • Adding this to every small task in an agent run. Anthropic says Sonnet 5.5 at low and medium effort can already check in more than you want.

Sources

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

Related

Beyond the course