How to Work with Eve to Create Chat Agent Instructions

Last updated: September 30, 2026

Overview

Agent instructions do not have to be written from scratch. Describing the task in plain language and asking Eve to convert it into instructions produces a working draft you can correct, which is easier than writing from a blank field.

This works in both directions. If you know what you want, describe it and have Eve draft the instructions. If you do not, ask Eve which agents would help based on what is in your matters, and start from the suggestions. Both are covered below.

Before You Start

Turn a Description into Instructions

1. Describe the task in plain language. Do not polish it. Say what you do, why, and what you want out of it, the way you would explain it to a new associate:

"Every time I get a new PI case I go through the medical records and build a timeline of every treatment: date, provider, what they did, what it cost. I need it chronological. I always end up hunting for gaps in treatment because that's what the adjuster attacks. And I don't want the billing summaries mixed in, just actual treatment."

That is a rambling description with the important material buried in it, and it is enough. Bullet points are not required. The substance is in how you would actually explain it.

If you have already done the task with Eve once, better: run the prompt in a normal conversation, refine until the output is right, then ask for the working version to be converted into agent instructions.

2. Ask for agent instructions. Any of these work:

"Make me a chat agent I can use on any matter that does what I just described."

"Turn that into instructions for a chat agent. It needs to work on any matter, not just this one."

"Write chat agent instructions for this. Break it into sequential steps if it needs more than one."

The critical phrase is "any matter". Eve is looking at the current case while writing, and without that instruction the output bakes in specifics: this plaintiff, this defendant, this set of records. Those details make the agent useless the moment it runs somewhere else. Saying "any matter" up front prevents the most common failure in the process.

3. Read the draft critically. Check four things:

  • Is it portable? Scan for anything specific to the current case: party names, dates, dollar figures, document titles that only exist here. Those need to become general instructions.

  • Does it say what to produce? Format, structure and length should be explicit.

  • Does it include your exclusions? In the example above, "no billing summaries" was easy to say and easy to drop. Exclusions get lost more often than inclusions.

  • Is it broken into steps if it should be? Multi-stage work runs better as sequential prompts than one dense block.

4. Pressure-test it before saving. This is the cheapest quality check available:

"Would these instructions work on a matter with no deposition transcripts?"

"What's missing from these instructions that would make the output inconsistent between cases?"

"Where would these instructions produce a bad result?"

The answer usually names a gap. Fix it in the thread and ask again.

5. Save the instructions as an agent. Two ways:

  • Copy and paste. Copy the instructions, go to Agents then Create Agent, select Chat Agent, and paste them into the instructions field. If the work was broken into sequential steps, paste each step into its own message using + New Message.

  • Magic wand. Paste the instructions back into the chat as your own message, then click the magic wand icon beneath it. Agent creation opens with the text loaded.

From there, name it, tag it, set visibility and save.

Asking Which Agents Would Help

Eve already knows your matters, so if you're unsure of where to start, you can use this to find agents that could help progress your real cases.

6. Ask for suggestions. Open a matter with real documents in it and ask:

"Based on what you know about this matter, what chat agents would help me?"

"What tasks am I probably repeating across my employment cases that a chat agent could handle?"

"I'm a solo PI attorney doing a lot of litigation. What three chat agents would save me the most time?"

The list comes back with reasoning attached.

7. Filter for what is actually repeatable. The test is whether you would run it on more than a handful of matters. A task done once a quarter is a prompt; a task done on every new intake is an agent. "Of those, which would I use most often across my caseload?" sorts the list.

8. Pick one and go back to the first workflow. Ask for the instructions, remembering the "any matter" phrasing.

Common Questions

  • Does the task have to be figured out before starting? No. A vague description produces a first draft, and reacting to a draft is easier than writing from scratch. If the description is too thin, Eve asks for what is missing.

  • Does this work for drafting and discovery response agents? No. Those need examples of past versions of the document. See 📄 How to Create and Use Drafting Agents in Eve.

  • Does this work for communication agents? Yes, same method. Name the type in the ask, such as "make me a communication agent for client status updates", so the instructions match what that agent type produces.

  • What if the instructions come back much longer than expected? Length is not a problem in itself; vagueness is. If it reads as padded, ask to tighten it: "cut this down to the instructions that actually change the output."

  • Can Eve write the name and description too? Yes, and both are auto-generated when the agent is saved. Either way, check that the description says which documents to select before running. That is the detail auto-generation misses most often.