Describing every row is not a narrative
Automated report writing usually converts a table into sentences. Twenty rows become twenty clauses. The reader learns nothing they could not have read faster in the table. The exercise only added length.
A narrative selects three movements worth discussing, in an order that makes an argument. The rest stays in the table where it belongs.
Selection requires knowing what the reader decides with. A figure that moved sharply and changes nothing is less newsworthy than a small move in the number a budget depends on.
Ordering is where the argument lives. Leading with the largest number is the easy default. Leading with the number that changes what someone does this week is the useful choice. Those two numbers are rarely the same figure.
How the agent is built in Agent Studio
Narrative standards live in the system prompt: how many movements a report may discuss, the requirement to state the context of any variance alongside any figure, and the instruction that a large movement within normal variance is not news.
Prompt skills carry your audience: what this reader decides, what they already know, and which vocabulary to avoid. That is what separates a broad narrative from a team narrative written from the same table.
A flow supplies the material with the arithmetic already done. A Math node computes variance and deltas before the agent sees anything, because a model reasoning about numbers it also calculated is two risks stacked.
Agent Chat hands over the computed table. The agent returns the selected movements, their order, and the reason for each selection.
A Generate Text node in the flow writes the prose from that selection. When the report is short, the agent writes the prose directly.
Google Sheets logs what was selected and what was skipped. A stakeholder asking why something was not mentioned is answerable from that log, and the question comes up more than you would expect.
What this agent is built from
- System prompt: how many movements may be discussed, and that a big move within variance is not news.
- Prompt skills: the audience, including what they decide, what they know, and what vocabulary to avoid.
- Math (Util): computes variance and deltas in the flow before the agent sees the table.
- Agent Chat (Util): hands the computed table over for selection and ordering.
- Generate Text (AI Core): writes the prose from the agent's selection.
- Google Sheets (Integration): logs what was selected and skipped, which stakeholders ask about.
- When an ActionFlow is enough: If you need a templated summary with slotted figures, build the workflow. That is fine for an operational report nobody argues with.
Frequently asked questions
Start building AI workflows
Create a free account, open a template or a blank canvas, and run your first ActionFlow.
Newsletter
Get product updates
New nodes, agents, and product notes. We send mail only when we have something worth opening.