Agent Atlas

Atlas · agents

Asking well

Your request is most of what the model knows about your task. Say what you want, why, and what done looks like.

Established practiceDepth 1 of 4Reviewed 2026-11-08

See it working: Plate I

In plain words

The model knows only what you tell it and what the harness loads. Anthropic suggests treating it like a brilliant new employee who does not yet know how you work.

How it works

Be specific about the result you want. Say why it matters and who it is for. In an agent, your request joins the instruction files and tool descriptions in the context, so standing rules belong in those files and the request can stay about this task.

Going deeper

Spoken requests work the same way once they become text. Some voice tools transcribe speech before the model sees it, so a misheard word becomes part of the request. On Plate I, change the request and compare what the agent does.

Sources

  1. Prompting best practices Anthropic
    • Think of Claude as a brilliant but new employee who lacks context on your norms and workflows.
    • Being specific about the output you want can help enhance results.
    • Providing context or motivation behind your instructions, such as explaining to Claude why such behavior is important, can give Claude more information about your goals
    • If you want Claude to do more than the task strictly requires, explicitly request it rather than relying on the model to infer this from vague prompts.
    • Providing context or motivation behind your instructions, such as explaining to Claude why such behavior is important, can give Claude more information about your goals and help it deliver more targeted responses.
    • If you say “can you suggest some changes,” Claude will sometimes provide suggestions rather than implementing them, even if making changes might be what you intended.
    • By default, implement changes rather than only suggesting them.
    • if you want the model to be more hesitant by default, less prone to starting implementations immediately, and only take action if requested, you can steer this behavior with a prompt
    • Claude Opus 4.5 and Claude Opus 4.6 have a tendency to overengineer by creating extra files, adding unnecessary abstractions, or building in flexibility that wasn’t requested.
    • A bug fix doesn't need surrounding code cleaned up.
    • Claude can sometimes focus too heavily on making tests pass at the expense of more general solutions
    • Tests are there to verify correctness, not to define the solution.
    • Otherwise, Claude might sometimes naturally attempt to bring its work to a close as it approaches the context limit.
    • The formatting style used in your prompt might influence Claude’s response style.
    • removing markdown from your prompt can reduce the volume of markdown in the output.
    • Never speculate about code you have not opened.
  2. Speech to text OpenAI
    • Send the audio file to /v1/audio/transcriptions

Quotes were checked against the live pages on 2026-11-14.