Asteria Docs
Agents

Tutorial: your first agent

Build a working agent end to end in about twenty minutes, using a job you actually have.

We are going to build Brand Voice Check, an agent that reviews copy against your style guide. Swap in your own subject as you go: the shape works for most jobs.

By the end you will have an agent you can call with @, and one you could share with your team.

Before you start

Have one document to hand that describes how your organisation writes: a style guide, a tone of voice page, even a good example of copy you were happy with. A single page is enough.

Put the reference material somewhere the agent can reach

If it is one or two pages, you will attach it directly to the agent later. Nothing to do now.

If it is a real pile (a style guide plus past campaigns plus product messaging), make a collection first and put them in it. Call it Brand Guidelines.

Create the agent

Agents in the sidebar, then Create Agent.

Name: Brand Voice Check

Description: Reviews copy against our tone of voice and flags anything off-brand.

Write the instructions

This is the part worth twelve of your twenty minutes. Adapt:

You check written copy against our brand voice for a B2B fintech
selling to finance directors at mid-sized firms.

How we write:
- Confident and plain. Short sentences. Concrete nouns.
- We explain, we do not sell. No superlatives.
- British English.

Never use: revolutionary, seamless, game-changing, unlock,
best-in-class, effortless.

When you review a piece of copy:
1. Give the corrected version first, ready to paste.
2. Then list what you changed and why, one line each.
3. Then flag separately anything that is a factual claim we
   would need to evidence. Do not rewrite those. Ask about them.

If the copy is already on brand, say so plainly rather than
inventing changes to justify yourself.

If you are unsure whether something breaks a rule, say you are
unsure and explain the tension. Do not silently pick a side.

The last two paragraphs are the ones that make it useful. Without them an agent will always find something to change, because that is what it thinks you asked for.

Give it the reference material

Scroll to Knowledge.

A collection: under Pinned Collections, pin Brand Guidelines. The agent now searches it automatically, with nobody naming it.

A page or two: under Uploaded Files, upload the style guide. Its content is in front of the agent on every turn.

Do not do both with the same document. Pick the one that fits the size.

Decide about tools

For this agent, probably none. It reads copy and gives an opinion.

Resist adding web search "just in case". A tool the job does not need is a way for the agent to wander off, and it makes every answer slower.

Add suggested prompts

Under Suggested Prompts, add two or three:

  • Check this landing page headline and subhead.
  • Is this email too salesy?
  • Rewrite this paragraph in our voice.

Skip this and a colleague opening your agent has to guess.

Save, then test it properly

Done to save. Now use the Test panel on the right, and do not test the easy case.

Give it copy that is already fine. Does it say so, or does it invent changes? If it invents, strengthen that instruction.

Give it copy with a factual claim. "Our platform cuts reconciliation time by 90%." Does it flag the claim rather than polishing it?

Give it something off-topic. "Write me a project plan." Does it decline sensibly?

Adjust the instructions and test again. Two or three rounds is normal, and this loop is the actual work of building an agent.

Use it

Back in the chat, type @ and pick Brand Voice Check, or click it in the sidebar under My Agents. Star it to pin it to the top.

Making it better over the next fortnight

An agent is not finished when you save it. The useful loop:

When it gets something wrong, fix the instructions, not the message. Rewording your request patches one conversation. Editing the brief fixes every future one.

Add the exceptions as you meet them. The first time it flags something it should not have, add a line saying so.

Keep it short. Instructions that sprawl to 4000 characters of special cases get worse, not better: the important rules drown. If it is getting long, the job may be two agents, or the specifics may belong in a skill.

When you are ready to share it

Give it to a colleague or two first and watch what they type. They will ask things you never would, and that is the cheapest feedback available. Then see Sharing and the marketplace.

Next

  • Skills for procedures you want in any conversation, not just this agent's.
  • Triggers to run an agent on a schedule.
  • Workflows for multi-step jobs that run without you.

On this page