# Write better prompts

Ask specific questions, scope them, and plan big tasks.

The agent reads your sources before answering, so the quality of the answer depends mostly on how clearly you describe what you want and where to look.

## Be specific [#be-specific]

| Instead of               | Ask                                                                              |
| ------------------------ | -------------------------------------------------------------------------------- |
| Tell me about this paper | What statistical methods did the authors use, and what were their limitations?   |
| Summarize my research    | What are the three most common themes across the papers in my climate folder?    |
| Papers on cancer         | Papers on PD-L1 inhibitors in triple-negative breast cancer published after 2022 |

Say what you will do with the answer, such as a table for a systematic review or a paragraph for a methods section, and the format you want back.

## Scope the question [#scope-the-question]

* Type `@` and pick a file or folder to point the agent at the sources you care about. See [Sources and @mentions](/guides/sources).
* Turn off web search when the answer must come from your documents.
* Select a passage in a PDF and send it to chat, rather than describing where it is.

## Plan big tasks [#plan-big-tasks]

For multi-step work such as a systematic review, type `@Plan` and ask the agent to outline its approach first. You can correct the plan before it starts, which catches misunderstandings early.

## Reuse what works [#reuse-what-works]

To save a prompt as a Skill, set rules for a folder, or have the agent remember a preference, see [Skills and instructions](/guides/customize).

## Control length [#control-length]

Ask for the length you want in the message itself, such as "a brief summary" or "a detailed analysis".

## Learn from examples [#learn-from-examples]

Each [use case](/use-cases/onboarding-library-setup) includes a full prompt you can copy and adapt, plus follow-ups.