Summary
In this guide we will give an overview on how to best talk to Eve as you get key work done throughout your case. This article will share best practices and advice on how to get the most out of your messages to Eve.
Before we get started
Instructions are for high-level information about a user, their firm, preferences and any pertinent background information they’d like to provide to Eve when working with Eve on a matter. Including instructions in your Session will help focus and create better quality answers in your messages to Eve.
Read more about instructions here.
Writing Messages to Eve, what you need to know
What are the messages I send to Eve and what should be included in them?
Messages are how you actually tell Eve to do work on different tasks. In technical jargon, these can be thought of as the normal Prompts sent to Eve inside a conversation.
Eve has high-baseline performance out of the box, but “prompt engineering” the messages sent to Eve can help increase her performance further for more advanced use cases..
How can Eve’s responses be refined?
It is extremely easy for a user to refine a response from Eve. To do so, simply tell Eve how you would like the response refined.
💬 Example Message: Please make that shorter and more concise. Focus only on the PAGA claims, not the FEHA ones.
“Prompt Engineering” - Optimizing your messages to Eve
As you work with Eve on your casework, you might find that you’re not getting a great response on the first “ask”. In this section we will share best practices and policies to employ as you try to optimize the answers Eve gives.
Here are some scenarios where these best practices may be particularly helpful:
You are having difficulty getting Eve to respond with high quality responses
You are creating a playbook for a common workflow from scratch
You are working on a complex or advanced use case
Note that this does not mean every message sent to Eve needs to follow these guidelines. Eve will respond to many messages quite well without a need for more complex messages - frequently called “prompt engineering”.
ℹ️ Note: The sequences of steps described below will walk through the example of using Eve to help evaluate a case: we will work from a simple starting point to a more thorough, well crafted sequence of messages, or prompts, which can be saved as a playbook for re-use at a later date.
Our example is intended to illustrate some different prompting best practices. Keep in mind that prompting is not an exact science, and is more of an art.
1. Provide detailed context and guidance
Generally, the more context, and guidance you provide to Eve on what you’d like her to do, the better the results will be. It can be helpful to include the following pieces of information in a message:
Task description: Describe the task you would like Eve to work on
Steps to follow: Provide Eve with steps to follow while working on the task
Here is a simple starting prompt for evaluating a case:
💬 Example Message: Please evaluate this case
This may lead to ok results, but we could see even better quality by re-structuring to the following:
💬 Example Message: We want to evaluate a potential employment law case in the state of California.
Please follow these steps:
Create a summary of the case
Identify potential claims
List out the key facts of the case in a timeline
List the witnesses and evidence
You will notice that the results after a prompt with more information about the context of your case are more tailored to the output we’d like to receive from Eve.
2. Break large tasks into smaller chunks
Eve can get overloaded when we ask her to do too many different things in one message. Note also that Eve is only able to generate responses that are around 2,000 words long.
It can be helpful to split more complex messages into multiple individual messages:
Case Summary
💬 Example Message: We are going to evaluate a potential employment law case in the state of California. We are considering representing the plaintiff in the case. I am going to give you instructions to follow for each step of the evaluation.
Please start by creating a short summary of the case that is roughly 100 words.
Notice how we maintained the pattern for our message: first a task description, then the steps to follow.
Now we can move onto the next step:
Claim Analysis
💬 Example Message: Now I’d like you to evaluate what potential claims this case has.
Please evaluate the case for each of these different types of claims:
Wage & Hour
Wrongful Termination / Retaliation
Discrimination / Harassment
Disability
And here are the remaining two messages:
Timeline creation
💬 Example Message: Now we are going to create a timeline of the key facts in the case.
Please follow these steps:
Read through the documents thoroughly as we don’t want to miss any key events.
Create a timeline with each of these key facts
Evidence / Witnesses
💬 Example Message: Now we are going to finish by evaluating the evidence and witnesses.
Please follow these steps:
List out all of the evidence
List any witnesses to the case
Notice how now we can be much more specific around which claims to look for. These messages can, of course, be modified for your use cases and case details.
3. Let Eve think before answering
For complex tasks, that may require a bit more reasoning by Eve, it can be helpful to have Eve think through responses step by step. We’ll apply this change to the Claim Analysis step as it requires Eve to make legal judgements:
💬 Example Message: Now I’d like you to evaluate what potential claims this case has. Please think through your reasoning step by step before providing your final answer.
Please evaluate the case for each of these different types of claims:
Wage & Hour
Wrongful Termination / Retaliation
Discrimination / Harassment
Disability
4. Standardize the format Eve responds with
Without providing guidance, Eve will choose the format to respond with on her own each time. In some cases, users may want consistent results when working with Eve on a specific task.
To account for this, you can provide Eve with details about the format she should respond with. It is often a good idea to do this at the end of your message. We’ll apply this to the Claim Analysis and Evidence / Witness message:
💬 Example Message: Now I’d like you to evaluate what potential claims this case has. Please think through your reasoning step by step before providing your final answer.
Please evaluate the case for each of these different types of claims:
Wage & Hour
Wrongful Termination / Retaliation
Discrimination / Harassment
Disability
Please respond with a section header for each type of claim, and a bullet point list of the potential claims of that type, with a description of each claim.
💬 Example Message: Now we are going to finish by evaluating the evidence and witnesses.
Please follow these steps:
List out all of the evidence
List any witnesses to the case
Please include a separate section for both the evidence and the witnesses.
Using tables and code blocks
Eve can respond with embedded UI elements in her answers, which can help make them more consumable. Two examples of these are tables and code blocks:
To generate output like this, you can tell Eve by saying “Put your response into a table” or “Place that in a code block”. These embedded elements also enable different actions including:
Save to Matter
Download as file (csv for table, docx for code blocks)
Copy
Here is an updated Timeline Creation message that outputs as a table:
💬 Example Message: Now we are going to create a timeline of the key facts in the case.
Please follow these steps:
Read through the documents thoroughly as we don’t want to miss any key events.
Create a timeline with each of these key facts
Please place your timeline into a table with the following columns:
Date: The date the key fact occurred, formatted like Month DD, YYYY
Event: A short description of the fact
And here is an example where the Claim Analysis is placed into a code block:
💬 Example Message: Now I’d like you to evaluate what potential claims this case has. Please think through your reasoning step by step before providing your final answer.
Please evaluate the case for each of these different types of claims:
Wage & Hour
Wrongful Termination / Retaliation
Discrimination / Harassment
Disability
Please respond with a section header for each type of claim, and a bullet point list of the potential claims of that type, with a description of each claim. Place this all into a code block.
In Conclusion
We started with a simple instruction to Eve for performing case evaluation:
💬 Example Message: Please evaluate this case
This works reasonably well, but we’ve demonstrated how we can significantly improve the results using the above best practices. As a rule of thumb, for complex prompts, you can follow this structure:
Task description: Describe the task you would like Eve to work on
Steps to follow: Provide Eve with steps to follow while working on the task
Formatting: Tell Eve how you would like your response formatted.
Here is the final sequence of messages for our case evaluation:
Case Summary
💬 Example Message: We are going to evaluate a potential employment law case in the state of California. We are considering representing the plaintiff in the case. I am going to give you instructions to follow for each step of the evaluation.
Please start by creating a short summary of the case that is roughly 100 words.
Timeline Creation
💬 Example Message: Now we are going to create a timeline of the key facts in the case.
Please follow these steps:
Read through the documents thoroughly as we don’t want to miss any key events.
Create a timeline with each of these key facts
Please place your timeline into a table with the following columns:
Date: The date the key fact occurred, formatted like Month DD, YYYY
Event: A short description of the fact
Claim Analysis
💬 Example Message: Now I’d like you to evaluate what potential claims this case has. Please think through your reasoning step by step before providing your final answer.
Please evaluate the case for each of these different types of claims:
Wage & Hour
Wrongful Termination / Retaliation
Discrimination / Harassment
Disability
Please respond with a section header for each type of claim, and a bullet point list of the potential claims of that type, with a description of each claim. Place this all into a code block.
Evidence / Witnesses
💬 Example Message: Now we are going to finish by evaluating the evidence and witnesses.
Please follow these steps:
List out all of the evidence
List any witnesses to the case
Please include a separate section for both the evidence and the witnesses. Place both in a table inside a markdown code block.
Immediate Next Steps
💬 Example Message: I’d like you to create a flowchart that outlines the immediate next steps we should take for this case after our evaluation.
Please include things like:
Client correspondence
Evidence to collect
Letters / Notices to file
Please create your flowchart using mermaid. Make it a horizontal flowchart and use colors for different types of actions. Only include a max of 10 next steps.
These messages can now easily be saved as a Playbook for re-use the next time we want to evaluate a case.
Saving Messages for Future Use - Playbooks
What are playbooks?
Playbooks are saved prompts that can be easily reused. As you are working with Eve, if you craft an excellent prompt, or sequence of prompts, you can save them as a playbook and run them again later.
They are also a great way to start a session with Eve for a common use case. Running a playbook can make it easy for a user to quickly start a session for their specific use case.
To learn more about how to save a message or sequence of messages (create a playbook) read the article here.
ℹ️ Note: As a reminder, playbooks are typically not intended to get a piece of work to a 100% final state. They are a great way to have Eve create a standardized starting point, at 80-90% completion.
From this starting point, the user can either instruct Eve to refine further, or manually refine themselves to get to the 100% state.
Using existing playbooks - editing for your use case
Eve comes with a number of playbooks for common Employment Law use cases. They are a great starting point for these use cases, and users should see good results out of the box.
That being said, every firm is different and users may want to customize and tune for the way they do their work. The out of box playbooks and instructions cannot be edited by an end user, but the messages can be copied by a user for use in their own custom playbook. To do so, users can follow these steps:
Run the out of box playbook
Create a new playbook by clicking the Teach Eve button on the message sent to Eve:
The user can add as many messages from the out-of-box playbook to their new playbook as they would like
If you have any questions about talking to Eve, don't hesitate to reach out to our support team at support@eve.legal.