# How to phrase a task

The working formula: **what to take → what to do → in what form to deliver**.

A short request looks faster but usually costs more, because clarifications follow. A detailed one closes in one or two turns.

## Before and after

| Instead of | Write |
|---|---|
| Analyse sales | Calculate revenue and margin per product for June, compare with May and show the 10 biggest drops. Deliver as a spreadsheet, one row per product |
| Evaluate the manager | Evaluate the manager's calls from last week against five criteria, each scored 1–10, with examples of good and weak calls. Deliver as a document |
| Make a deck | Make 6 slides: one problem per slide, photo, cause and next step. No statistics, no intermediate files |
| Check the mailbox | Collect emails from the last 3 days that nobody answered and tell me who needs a reply today |

## What is almost always worth adding

- **Period** — "for June", "from the 1st to the 14th", "last week".
- **Breakdown** — per product, per deal, per employee, per warehouse.
- **Output format** — spreadsheet, document, deck, short message, text ready to paste into a messenger.
- **What not to include** — "no intermediate files", "no general recommendations", "numbers only".

## An example that needed no clarification

<TryPrompt prompt="Build a report from @mango_office for 1–14 June for the “Support” group: total inbound, answered and missed calls; separately — called back, was called back, unreachable, spam; plus the average call-back time per day." />

The service, period, group and fields are all named, so the report arrives on the first try.

## Corrections: pointed, not "redo it"

When you get a draft, do not restate the task. Say what exactly to change:

- "Keep only the overdue deliveries."
- "Different signature format — like in the previous letter."
- "Add a cost column and recalculate the margin."

The agent remembers what it has already assembled and redoes only that part. "Start over" is the slowest route: it pulls all the data again.

**If the answer went completely off**
Starting a [new chat](/help/chats) with a sharper brief is faster than arguing in the old one, where the accumulated context keeps pulling the agent back to its earlier reading of the task.

## Effort

Next to the input box there is an **Effort** switch:

- **Normal** — quick answers, reformatting text, writing a letter.
- **Medium** — most working tasks.
- **High** — analytics, legal reviews, complex calculations. Slower, but digs deeper.

For a long brief, **voice input** helps: dictating the details is faster than typing them.