> ## Documentation Index
> Fetch the complete documentation index at: https://docs.open-hive.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Write a useful task brief

> Specify an outcome, sources, output format, limits, and review points so the result is easier to trust.

A useful task brief tells an agent what success looks like and how you will judge the result. You do not need to design the agent team yourself.

## Include five things

| Part                        | Example                                                                   |
| --------------------------- | ------------------------------------------------------------------------- |
| **Outcome**                 | Identify ten companies that could buy our onboarding software.            |
| **Context**                 | We sell to US software companies with an in-house customer-success team.  |
| **Sources and access**      | Use public company pages and the attached customer profile.               |
| **Deliverable**             | Save a table with website, fit evidence, source URL, score, and unknowns. |
| **Limits and review point** | Research only. Stop after ten results; do not contact anyone.             |

## Copy this structure

```text theme={null}
Goal:
[The result I need and why.]

Context:
[Audience, company, existing work, definitions, and exclusions.]

Inputs:
[Files, links, connected account, and sources to use.]

Output:
[Format, required fields, where to save it, and how I will assess it.]

Limits:
[Batch size, time period, permitted actions, and things to avoid.]

Finish:
[When to stop, what to summarize, and what needs my review.]
```

## Separate research, drafting, and sending

Each is a different instruction. “Draft a reply” and “send a reply” have different consequences. Name the intended account and recipient when authorizing a send, and give a concrete stopping point.

A request for approval in a prompt is an instruction to the agent, not a universal technical lock on all tools. Review the full template and the tools you enable before running it.

## Make recurring tasks safe to repeat

A repeated task should explain how to recognize work already completed:

```text theme={null}
Check the three specified company news pages for changes since the
last successful run. Save new items with source URL and first-seen date.
Skip URLs already recorded. If nothing changed, report that briefly.
Do not send email or update external systems.
```

Test once before adding a [schedule](/guides/scheduling). Avoid an instruction such as “keep doing this” without defining the interval and what counts as new work.

## Correct the result precisely

Instead of “this is wrong,” point to an example and a rule:

```text theme={null}
The third company sells to agencies, not to our target audience.
Exclude agency-focused vendors and recheck the list using that rule.
Preserve the rows that still qualify.
```

For context you want reused, consider [memory](/concepts/memory). For a repeatable procedure, use a [skill](/concepts/skills). A task-specific instruction usually belongs in that colony's chat.