Skip to main content

Write run instructions for agent Private Preview

TL;DR

Learn how to write run instructions that tell a context agent template exactly which terms to extract from your documentation.

Connect

Run instructions tell an agent what to extract from the sources it can reach. They are the highest-leverage field in the setup flow: most first-time agents underperform because of what's written here, not because of the model.

info

Run instructions are per-agent and are sent with every run, including scheduled ones. They're different from custom instructions, which are org-level and apply to collection enrichment across your whole workspace.

Prerequisites​

Before you begin, make sure you have:

  • Workspace admin or governance admin access
  • An agent created from a template, or the create-agent wizard open at the Guidance step. See Create an agent from a template

Where to find run instructions​

You can set them during setup or change them at any time afterwards.

  • During setup, on the Guidance step of the create-agent wizard.
  • After setup, on the agent's Settings tab.

Changing run instructions takes effect on the next run and doesn't create a new version of the agent.

What to write​

You have 1000 characters, which rewards precision over completeness. A pattern that works:

Extract [what kind of concept] from [which part of the source].
Focus on [the specific domain or subject area].
Ignore [what you do not want].
For each term, [any convention you need].

An example​

Too vague:

Create glossary terms from our docs.

This gives the agent no way to decide what counts as a term worth creating, so it returns whatever it finds first.

Specific enough to act on:

Extract business terms from the settlement and chargeback runbooks only. Focus on terms used by the operations team in daily processing: status codes, exception types, and settlement windows. Ignore infrastructure and deployment pages. Keep definitions to two sentences and name the source system where the term originates.

What to include​

Say what to ignore​

Most sources contain more noise than signal. Naming the noise usually helps more than naming the signal, because it removes whole sections from consideration rather than asking the agent to rank within them.

Example:

Ignore release notes, meeting notes, and anything under the Archive space.

Set length of definition​

Without an instruction, definitions inherit whatever length the source happened to use, which reads inconsistently across a glossary.

Example:

Keep each definition to two sentences: what the term means, then how it is used.

Name your audience​

Who the definition is written for changes the output substantially.

Example:

Write for a business analyst who does not know our internal acronyms. Spell out every abbreviation on first use.

Point at subject area, not just source​

Scope narrows where the agent looks. Run instructions narrow what it extracts once it's there. Both help.

Example:

Focus on credit risk terminology. Skip anything about payments or servicing even where the two appear on the same page.

Tips​

  • Be specific, not exhaustive. Three precise sentences outperform a long paragraph. The agent reads your instructions on every run, so keep them signal-dense.
  • Preview after every change. A sample run takes a minute or two and publishes nothing. It's the cheapest way to find out whether an instruction landed.
  • Change one thing at a time. If you rewrite the scope and the instructions together and the output gets worse, you can't tell which change caused it.
  • Don't include sensitive information. Instructions are used in AI prompts. Avoid credentials, internal URLs, and confidential business detail.

Need help​

If you need assistance after reviewing this page, contact Atlan support.

See also​