Give your agent its job description

Why this matters

Every agent in Auxot has a job description, and the sharper that description, the better the agent’s answers sound like yours. “Helps with marketing” gives the agent no voice, no constraints, no sense of what good marketing looks like for your business. “Drafts posts and emails using our brand voice. Analyzes campaign performance. Suggests content ideas. Never makes pricing claims without checking.” tells the agent what work it takes on, in what voice, with what limits, so on marketing questions you get marketing help that sounds like yours.

In Auxot, an agent’s job description is four parts: its name, its Description, the context files it draws on, and the tools it can reach for. Edit any of them and the next message reflects the change: no Save button, no deployment step. Today, sharpen one description. The next time you use that agent, it stops giving you generic answers and starts giving you yours.

Quick start

1. Sign in: open Auxot in your browser and log in.
2. Open Agents settings: click Settings in the left menu, then Agents.
3. Click the agent you want to edit: any of your custom agents. (The Admin Agent is locked since it ships with Auxot: you can’t edit its description, but you can build your own agents that look up to its standard.)
4. Update the Description field: a short summary of what this agent is for. Auxot saves automatically as you type.
5. Scroll down and attach context files: these are where the real instructions live. Click Add a context file and pick from what you’ve already created.

Done? Your agent now has a clear job description. The next conversation should already feel different.

The agent can do that?

Now that your agent has a job description, pull the Admin Agent into the loop the same way you would anywhere else: you open Chat, pick the Admin Agent, send a prompt first, and he answers from there. The three gray boxes below are those prompts. Each reply is copy-ready text you paste into your other agent’s chat when you’re ready.

1. Draft a first real piece of work (you ask; he answers with a paste-ready prompt)

Open chat with the Admin Agent and try this:

Look at my "[agent name]" — its Description and the context files I've attached. Design a first real assignment I can hand to it today: something specific, with a concrete deliverable, that would actually be useful to me this week. Write it as a prompt I can copy-paste straight into chat with that agent.

Why it’s non-obvious: Most people configure an agent and then try to invent a good first prompt from scratch. Run this in Admin Agent chat instead and his reply is built from the Description and context files you just set up: a concrete first assignment for that specific agent, ready to paste.

2. Have the Admin Agent design a recurring job your agent should own

My "[agent name]" needs a job it does every week without me having to remember to ask. Based on its Description and context files, what's a recurring task this agent should be running? Draft the Description to include this weekly item, as well as the initial prompt I'd give it. Tell me whether any context files would need to be added to make the output usable.

Why it’s non-obvious: Most people use agents one-off: prompt, answer, done. Paste the gray-box prompt above first; once you do, he reads what you configured about that agent, surfaces what ought to repeat weekly, bakes that into rewritten Description wording, and hands back a kickoff prompt. All in one reply. The rhythm is now on paper: each run still starts when someone kicks it off. You paste the kickoff yourself, or put that kickoff on a schedule you configured. Run an agent on a schedule is where a single recurring prompt lives on the clock (Settings → Agents → Scheduled Tasks). Run a workflow is where multi-step chains live without a sticky note. Combine with Define a tool policy when the recurring job reaches for tools beyond plain language.

3. Get a stretch-task prompt from the Admin Agent, then bring the answer back

Pick a hard assignment for my "[agent name]" — something at the edge of what its Description actually covers. Write me the prompt. After I've handed it to the agent and gotten an answer, I'll paste the answer back to you and we'll figure out together where the Description, context files, or tools need tightening.

Why it’s non-obvious: Start by sending the gray-box prompt. His reply is the stretch prompt for you to paste into your other agent’s chat; you run it there, bring the answer back into Admin Agent chat, and he helps you see what to tighten. Replaces “discover the gaps slowly over weeks” with “find them in one afternoon.” You ran the other agent; he supplied the test question and read what came back with you.

Go deeper

Why the Description field is so short

The Description in the agent edit page is short on purpose: it’s the brief summary that shows up wherever the agent is referenced (in the agent picker, in the agents list, and in audit logs). If it were a long block of text, those places would be cluttered or truncated.

The actual “long-form” instructions live in context files attached to the agent. That’s where you put the detailed do’s and don’ts, the example outputs, the policy text, anything that won’t fit in the brief Description. The Description is a short summary of what the agent does; the context files are where the detailed instructions live.

Troubleshooting: the agent isn’t following its description

• The Description is too vague. “Helps with marketing” gives the agent nothing to enforce. Tighten it: name a specific kind of help, name the constraints.
• The constraints aren’t in the right place. Hard rules (“never offer discounts”) work better in an attached context file with a heading like “Rules I must always follow.” The Description is a summary; context files are where rules live.
• The agent doesn’t have the context files it needs. Check the Context Files card on the agent page. Make sure the relevant ones are attached. Without the brand voice file, the agent can’t write in your brand voice.
• The agent has too many context files. Counterintuitively, attaching every file you have can dilute focus. Keep each agent’s files relevant to its job.

Variations & edge cases

• Built-in agents are locked. The Admin Agent’s description and configuration ship with Auxot: you can’t edit it. To get an Admin-Agent-shaped agent that you control, build a custom agent with similar context files attached.
• Auto-save means no Save button. Some users go looking for one. There isn’t. Type your changes; they save in the background within a second or two. A small toast confirms each save.
• Preferred AI model can be set per agent. The “Preferred inference provider” dropdown lets you pin a specific model to this agent, which is useful if you want a particular agent to always use Claude (or always use GPT-4, or always use your local GPU). Optional; defaults to your account’s default.
• Tools and context files are configured separately. Tools come from “tool policies” (Define a tool policy), and they’re attached the same way context files are, via a card on the agent page.
• Editing on Free tier: all agent editing works on Free tier the same way it does on paid tiers. The only thing Free tier can’t do is move agents between teams (since Free tier doesn’t have multiple teams).

• When you ask the Admin Agent to create an agent for you. In one go (after you confirm in chat), it can name the agent, write its Description, set its system prompt, pin a preferred AI model, attach context files, and configure its tool policy. Everything lands at once or nothing does.

Walkthrough

Step 1: Sign in

Open Auxot in your browser and sign in.

Step 2: Open Agents settings

Click Settings in the left menu, then Agents. You’ll see a list of every agent on your account: the Admin Agent at the top (it’s there by default), and any custom agents you or your team have created.

Step 3: Click the agent you want to edit

Click the row for the agent you want to give a job description to. The detail page opens, with editable fields for everything that makes the agent who it is.

A few things to know up front:

• The Admin Agent is locked. It ships with Auxot and its description is fixed. You can’t edit it, but you can take inspiration from how it’s set up, and you can build your own agents alongside it.
• Built-in tool count is read-only. The “tools” line tells you how many tools the agent can reach for; that’s set elsewhere (Define a tool policy).
• Auto-save is on by default. As you edit fields, Auxot saves in the background. You don’t need to hit Save: there isn’t one. (If you change your mind about an edit, just type the original value back in.)

Step 4: Update the Description field

The Description is a short summary of what this agent does. It’s brief on purpose: concise enough that anyone scanning the agents list understands what this agent does at a glance.

Three rules of thumb for writing good descriptions:

• Say what it’s for, not what it is.
• Name the constraints. If there are things the agent should never do, say them here: “Never offers discounts. Always offers a human handoff for technical issues.”
• Be specific. The specifics are what guide the agent’s voice toward yours.

Some examples that work, narrow scope and broad scope both:

• “Drafts customer support replies using our brand voice. Never offers discounts. Always offers a human handoff for technical issues.” (narrow scope, single recurring task)
• “Our marketing agent: drafts posts and emails, analyzes campaign performance, and suggests content ideas. Always uses our brand voice. Never makes pricing claims without checking.” (broad scope, multiple related tasks)
• “Summarizes weekly all-hands meetings into one paragraph each, focused on decisions made and action items assigned.” (narrow scope, single recurring task)
• “Our research analyst — pulls together briefings on prospects, summarizes industry trends, and flags news that affects our customers. Cites sources for every claim. Never speculates beyond what the source says.” (broad scope, multiple related tasks)

Both scopes are valid. What matters is naming the work, the voice, and the constraints, not how many tasks the agent juggles.

Auxot saves your changes as you type: there’s no Save button, and the field updates live.

Tip: Don’t try to write the description perfectly the first time. Write a rough version, talk to the agent, and refine the description based on what it gets wrong. Job descriptions are living documents.

Step 5: Attach the context files that do the real work

Below the Description, you’ll see a card titled Context Files. The page’s own helper text describes it well: “No context files yet. Add at least one so the agent has written guidance to follow.”

This is where the actual “instructions” live. The Description is a short summary of what the agent does; the context files are where the detailed do’s and don’ts go.

Click Add a context file. A picker appears showing every context file you’ve created (and any Org files your team has shared). Pick the ones that match what this agent does. For a customer support agent, that might be your brand voice file plus a customer FAQ. For a contract reviewer, it’d be your compliance rules file plus a few past contract examples.

You can attach multiple files. After you send your next chat message, the agent can pull from every file attached to it, whichever ones fit that question.

Tip: If you don’t have the right context file yet, head back to Add your first context file to add one, then come back here and attach it. Agents are easier to refine after they’ve got something to work with.

Nothing restarts behind the scenes: type a message in that agent’s chat and you’re exercising the new attachments right away. Test it by asking something where the file should clearly matter.

What’s next

• → Invite your first teammate. Once your agents are doing real work, bring others in to use them.
• → Create an agent from scratch. Same Description + context files + tools, but starting from a blank slate. Describe the job in Chat and he walks you through the rest; you still prompt first.
• → Define a tool policy. Give your agents real-world capabilities (search, code, internal APIs) on top of the context files.
• → Run a workflow. When recurrence should become a repeatable chain with a manual start, a webhook intake, or a scheduled trigger, not another sticky note above your keyboard.

Reference

• Pages in Auxot: Settings → Agents
• Built-in agents: locked (the Admin Agent)
• Save behavior: auto-saves as you type; no Save button
• Job description = Name + Description + Context Files + Tool Policies (composed across the agent’s detail page)
• See also: Add your first context file, Create an agent from scratch