Write good task instructions

The agent knows only what your brief says and what you give it access to. This guide covers what to put in a brief, how to break up big work, and how to steer and iterate once the task is running.

The parts of a good brief

A good brief answers six questions. You don’t need all six every time, but each one you skip is a guess the agent has to make.

Part	What to say	Example
Goal	What the result is for	“So the board can decide whether to open a second warehouse”
Reader	Who reads it and what they already know	“The Harbor & Pine board. They know the business, not the logistics terms.”
Output	The kind of file and its shape	“A 10-slide deck”
Sources	Where the facts come from	“Use the attached Q3 shipping report and the 2026 budget in Drive”
Example	What good looks like	“Match the structure of last quarter’s deck, attached”
Constraints	Rules the result must follow	“No more than 30 words per slide. Every number cites its source.”

Before and after

A short brief leaves most of the decisions to the agent:

Make a deck about the warehouse.

The same request with the six parts filled in:

Make a 10-slide deck for the Harbor & Pine board on whether to open a second warehouse in Q2. The board knows the business but not logistics terms, so explain any you use. Use the attached Q3 shipping report and the 2026 budget sheet in Drive. Follow the structure of last quarter’s board deck, attached. Keep each slide under 30 words, cite the source for every number, and end with one recommendation and its cost.

The second brief takes a minute longer to write. It also tells you how to judge the result: you can check the slide count, the word count and the citations yourself.

Say what to make

Pick the kind of result before you write. On a workspace home, the Make and Build badges under the title set the result type:

Make is for documents (Presentation, Document, Spreadsheet, Chart) and media (Image, Diagram, Video).
Build is for apps and tools (App, Dashboard, Website, Agent), automations (Scheduled and Monitor, plus Triggered where your organization has it turned on) and, if you have Evals access, Evaluation.

Picking one doesn’t send anything. It sets the type, and you still write the brief. If you’d rather not use the buttons, name the file type in the brief, for example “a .xlsx with one sheet per region”.

Give it the sources

The agent can only use what you attach, mention or give it access to. If a fact matters, point to where it lives.

Files: click + and choose Attach files, or paste a file into the composer.
Drive: click + and choose Drive, or paste a Drive link.
Earlier work: click + and pick from your recent tasks.
A file or skill by name: type @ and pick it.
Connected apps: choose which connectors the task may use under Connectors.

Say which source wins when they disagree, for example “If the budget sheet and the report differ, use the budget sheet.”

See Send a task for each step in the composer.

Show what good looks like

An example of finished work does more than a paragraph of description. Attach last quarter’s report, a memo you liked, or a short sample you wrote yourself, and say what to copy from it: the structure, the tone, the level of detail, or all three.

If the example contains things you don’t want repeated, say so: “Use the section order from the attached memo, but not its numbers.”

Write constraints the agent can check

Write rules that have a clear yes or no answer. The agent can check “Under 400 words” or “Every claim links to a source.” It can’t check “Make it concise” or “Make it look professional.”

Good constraints usually cover:

Length: a word count, slide count or page count.
Format: headings, tables, file type or naming.
Scope: what to leave out, such as “Only stores in the Northeast region.”
Actions: what the agent must not do, such as “Draft the email, but don’t send it.”

Break up big work

For a long or risky job, don’t send one huge brief and hope. Use one of the two task modes. Click + in the composer and turn on Plan mode or Goal mode. You can use one mode at a time.

Plan mode

With Plan mode on, the agent lays out its plan and checks it with you before it acts. Use it when a wrong first step is expensive: a change to many files, a long research job, or anything that writes to a connected app.

Read the plan as you would a colleague’s. Correct the order, cut steps you don’t need, and add the checks you care about before you agree.

Goal mode

With Goal mode on, your next message becomes the goal, and the agent keeps working until it reaches it. The goal stays pinned above the conversation. From there you can Pause goal, Resume goal, Complete goal or Clear goal.

Goal mode works best when you can describe “done” in a way anyone could check:

Before: “Clean up the customer list.”

After: “Every row in the attached customer list has a valid email, a country, and no duplicate company names. Put any row you can’t fix on a separate sheet called Needs review.”

Split by hand

You can also split a job into a series of tasks yourself: one for research, one for the draft, one for the final file. Attach the earlier task’s result to the next one with +. Each task stays short, and you can check the result at every step.

Steer while it runs

You don’t have to wait for the task to finish to change course.

Add an instruction. Type in the composer and send. The message waits under Queued messages, and the agent reads it at its next step.
Answer approvals. When an action is set to Ask first, the agent stops and asks before it acts. Answer quickly, because a task that’s waiting doesn’t progress.
Stop. If the task is going the wrong way, click Stop and send a correction.

Correct early. A one-line correction after the first step saves a full redo at the end. See Monitor and steer a task for each control.

Iterate on the result

When the result is close but not right, choose how to continue:

Situation	What to do
A few things need fixing	Send a follow-up in the same task, such as “Shorten slide 4 and add the cost to slide 9.”
You want to try a different direction and keep this one	Open ⋯ > Fork. It copies the whole conversation into a new task.
The brief itself was wrong	Start a new task with a better brief.

Follow-ups in the same task keep everything the agent has already read and made. A new task starts from nothing, so use it only when the old context would get in the way.

Make each follow-up specific. “Make it better” gives the agent nothing to act on. “The summary is too long. Cut it to three bullet points and lead with the cost” does.

Move standing instructions out of the brief

If you paste the same rules into every task, move them somewhere they apply on their own.

The rules describe	Put them in	Example
How to do one kind of job	A skill	“How we write board decks: 10 slides, one idea each, sources in the notes”
Who the agent is, what it may use, and how it behaves in every task	The agent’s Instructions	“You answer sales questions. Use @Product Docs first. If the answer isn’t there, say so.”

A skill can be attached to any task or added to any agent, and you can improve it over time. See Build and improve skills.

Write agent instructions

On an agent’s Setup tab, the Instructions field says what the agent does and how. Type @ in it to link a skill, file, connector or another agent.

The same rules apply as for a brief. State actions and conditions: what to do in which situation, and what not to do. “Be helpful and accurate” doesn’t change anything. “If a question is about pricing, use only the current price list, and say which version you used” does.

Keep the instructions about the agent’s job and limits. Put house style and repeatable steps in skills, so other agents can use them too. See Create an agent for an example.

Next steps

Send a task: every option in the composer.
Monitor and steer a task: queued messages, approvals, Stop and Fork.
Build and improve skills: turn standing instructions into a skill.
Create an agent: set up an agent’s instructions, skills and connectors.
Choose a model: pick a model for the job.