Tips & Best Practices

Get the most out of the Operator Agent — effective prompting, reading the agent's output, and knowing when to use the agent vs. a workflow.

Writing Effective Prompts

The agent responds to natural language, but how you phrase your request makes a real difference.

Be specific about the task:

Instead of...	Try...
"Process this file"	"Extract all line items from this PDF and create a CSV with columns for SKU, quantity, and unit price"
"Check this email"	"Read this email and draft a reply confirming we can deliver by the requested date"
"Analyse this data"	"Compare Q1 and Q2 revenue by product category and highlight anything that changed by more than 10%"

Provide context upfront:

• Upload all relevant files with your first message

• Mention which systems or data sources are involved

• Specify the output format you want (CSV, email draft, summary, etc.)

Iterate naturally:

The agent remembers everything in the current session. You can:

• Ask it to revise: "Change the date format to DD/MM/YYYY"

• Build on previous work: "Now do the same for the March data"

• Redirect: "Actually, focus on the top 10 by revenue instead"

Understanding the Agent's Output

Narration

While working, the agent shows short status messages describing each step. This is useful for:

• Knowing it's still working on longer tasks

• Catching mistakes early — if the narration shows it's headed in the wrong direction, you can interrupt

• Understanding the approach the agent took, which helps you refine future requests

Todo List

For complex tasks, the agent creates a visible todo list showing its plan. Items are checked off as they're completed, giving you a clear picture of progress.

Skill Cards

When the agent invokes a skill, a card appears in the chat. Skill cards show:

• The skill name

• Execution progress

• Results when complete

Cards appear immediately in a loading state and update in place — nothing shifts around or disappears.

When to Use the Agent vs. a Workflow

Use the agent when:

• The task is ad-hoc or one-off — you won't repeat it the same way every time

• The task needs judgement or reasoning — interpreting documents, drafting responses, comparing data

• You want to explore or investigate — "What's in this file?", "How does this compare to last month?"

• The task involves multiple steps that depend on what's found — the agent adapts as it goes

Use a workflow when:

• The process is repeatable and structured — it runs the same way every time

• You need scheduled or triggered execution — process every incoming email, run every morning at 9am

• Volume is high — processing hundreds of orders, invoices, or documents per day

• You need deterministic, auditable results — every run follows the exact same logic

Use both together:

• Use the agent to prototype and test a process, then convert it to a workflow once it's proven

• Use workflows for the high-volume automation and the agent for exception handling and edge cases

• Use skills to bridge the gap — a skill created in the agent can encode a process that runs reliably every time

Common Pitfalls

• Being too vague. The agent will try its best, but vague instructions lead to generic results. Specificity wins.

• Not uploading files upfront. If the agent needs data, give it the data in your first message rather than making it ask.

• Expecting workflow-level determinism. The agent makes decisions — it might take a different approach to the same task on a different day. If you need consistency, use a workflow.

• Ignoring the narration. The narration tells you what the agent is doing. If it's going wrong, catch it early and redirect.