> ## Documentation Index
> Fetch the complete documentation index at: https://docs.rundesert.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Writing a good prompt

> How to describe a job so the agent gets it right the first time.

The agent asks you questions when your description leaves gaps. That works, but
it costs you a round trip every time.

Answer those questions up front and the agent goes straight to building. This
page covers what to put in that first message.

One rule explains most of it. **You own what the job is. The agent owns how it
gets done.** So tell it the outcome you want, hand over the access it needs, and
let it pick the tools. You never have to say which API to call or where to store
anything.

## Say what you want out of it

This is the big one. Most weak prompts name a task and stop, which leaves the
agent guessing at what a good result looks like.

Compare these two.

<CodeGroup>
  ```text Weak theme={"system"}
  Summarize my code changes.
  ```

  ```text Strong theme={"system"}
  Summarize my code changes. I want to know the impact of the change, which
  downstream services it affects, and the main files that changed with an
  explanation of each one.
  ```
</CodeGroup>

Both ask for a summary. Only the second one tells the agent what the summary is
for, so only the second one can come back with the right summary. The first one
gets you a list of commits, which is technically correct and useless.

Try finishing your sentence with "so that I can." If you cannot finish it, you
have not said your intent yet.

## Logins and tokens

Ask yourself which service the job needs, then check its settings page for a
personal API token. Most services have one. GitHub, Stripe, Linear, Notion, and
Airtable all do.

Give the agent a token instead of your password whenever you can. A token does
not expire when you change your password, it never trips two-factor, and it lets
the agent skip driving a browser through a login screen, which is slower and
breaks more often.

Paste the token right in the chat when you have it. The agent saves it to a
locked file in your workspace and reads it only when a run needs it. It never
puts your token into the app itself.

<Note>
  Check our [Integrations](/integrations/overview) first. If we already support the
  service you need, connect it there instead of pasting a token. We manage those
  credentials for you, which makes them both safer and steadier: your long-lived
  credential never reaches your workspace, and we refresh it before it expires so
  your automation does not break one morning.
</Note>

## Correct it as it goes

You will not get everything into the first message, and you do not have to.
Watch the first few runs and say what is off.

> Too long. Three bullets, not ten.

> You marked that one urgent. It was not, because they are on the free plan.

The agent writes each correction into its memory in that workspace, so it carries
forward. See [It learns from your corrections](/#the-agent-remembers-your-corrections).

## Before you hit send

<Card title="Quick check" icon="list-check">
  * Did I say what the result is **for**, not just what to do?
  * Did I include the **token** or say which account to use?
</Card>