AI personalization

Give the agent standing instructions it follows on every new workflow and Workspace chat

AI personalization gives the agent standing instructions in Markdown. The agent reads them when it starts a new workflow or Workspace chat.

It’s especially useful for rules that apply across all your workflows, for example how your team names fields, which date format your warehouse expects, which sources are off limits, and who to ask when something is ambiguous.

​

Where to set it

Go to Settings > General and edit the AI personalization field, then click Save.

Anyone in the workspace can edit it. Renaming the workspace still requires an Admin or Owner.

The field takes Markdown and holds up to 8,000 characters.

​

When it takes effect

The text is read once, when a workflow’s agent session or a Workspace chat is created. Every workflow has one long-lived session, so:

• Workflows, real-time monitors, and Workspace chats you create after saving an edit follow the new text.
• Workflows that already exist keep the instructions their session started with. Reopening the chat resumes that session, so it does not pick up the edit. To apply new rules to an existing workflow, say them in the chat or put them in the workflow prompt.

Open Workspace chats keep their text too. Start a new chat to pick up the edit.

Personalization applies to the whole team, so the agent follows it for everyone. If you belong to several workspaces, each one has its own text.

​

What to write

Write it like onboarding notes for a new colleague. Short, concrete rules beat long prose.

## Output conventions
- Dates are ISO 8601 (YYYY-MM-DD), always UTC.
- Prices are numeric, no currency symbol, with a separate `currency` field (ISO 4217).
- Use snake_case for every field name.

## Our sources
- "The portal" means partners.acme-internal.com. It needs the login stored in the `portal_login` variable.
- Never extract personal contact details from job postings. Company-level fields only.

## How to work with us
- If a field is missing on more than half the rows, ask before guessing a fallback.
- Prefer failing a run loudly over delivering partial data.

A few more examples of instructions that pay off:

Goal	Instruction
Consistent naming	”Name the primary identifier external_id, never id.”
Regional defaults	”Unless told otherwise, extract the German (de-DE) version of a site.”
Compliance boundary	”Do not extract user reviews or any author names from review sites.”
Escalation rule	”If a source starts blocking us, ask before switching to a different source.”

​

What not to put here

• Secrets. Store credentials as variables or secrets and reference them by name.
• Instructions for one workflow. Those belong in that workflow’s prompt or in a template.
• Schema definitions. Define fields in the workflow schema, where they are validated per run.

​

Next steps

Agent chat

Build and debug workflows by chatting

Prompts

Write per-workflow extraction instructions

Templates

Share prompt and schema settings across workflows

Variables

Store reusable values and credentials