AI Guidelines
Large Language Model (LLM) tools (often referred to as "AI") provide general purpose capabilities relevant to work tasks including research, synthesis, writing, coding, troubleshooting. Pluribus encourages use of LLMs to help you be effective in your role. Do avoid liabilities of such tools, including “work slop” or violations of formal customer policies. Proper use of tools means you are empowered but accountable.
Here’s how to be successful in that…
- You have permission to use the tool. For customer information, we must have the rights to send that data to a given tool, even if we are operating outside of the government network. Consult the below sections on customer or internal environments. The rules may be different for documentation, code, or data.
- You are accountable for quality. You put your name on the work, no matter how it was produced. Do not simply prompt and then hand it off as “here’s what claude says”. Own it. Review, re-write, re-prompt, until you have something of value that you will stand behind.
- People make decisions. People provide direction. Tools are there to execute the tasks you direct. You are the authority, not the bot. Your value is in determining what we should do and how it should get done. In this, you are smarter than the bot. Use a tool as brainstorming partner. “What should I include in a plan? What am I missing in this plan?” You are accountable for decisions and direction.
- Use the tools to enhance your craft. Read the docs. Learn what makes a good LLM use case. Be thoughtful about where in your workflow to use tools. Offload drudgery. Retain ownership of your thoughts and intentions. Work at the right level of abstraction. Sub-par output happens for many reasons. The task may be a poor fit for LLMs. A skill might better guide the tool. Your prompt may be improved. You might directly edit output.
- Consider cost. Pay attention to token usage and cost. Better prompts and model selection can yield huge differences in efficiency.
Note: this summarizes and extends the formal policy in the Employee Handbook. Consult the handbook for additional detail.
Guidelines for Customer Environment Usage
We often operate either in a customer's technical environment or in the Pluribus environment. Some of us may switch between the two. The following order of operations applies, with the first items outranking the last.
- Government environment: You are using government-managed equipment, networks, and software. In this case, tools may be physically restricted (certain apps or URLs blocked) in addition to policy restrictions. In this case you must follow the customer policies and controls in place. Consult your program manager or account lead for guidance.
- Government information: You are using Pluribus equipment but have access to some government information, be it source code, files, or anything else. Your equipment might have tools available, but applying an AI tool to any of the government information must comply with any policy or contractual constraints. Consult your program manager or account lead for guidance.
- Pluribus information: You are using Pluribus equipment and information. In this case, you have the most flexibility to use tools, but there are still guidelines (immediately following).
Guidelines for Internal Environment Usage
When using Pluribus equipment and Pluribus information, the following restrictions apply. If you need further clarity, ask your manager and/or the helpdesk.
External Tools: You might make some use of tools that are not licensed by Pluribus. Perhaps you have a personal license or are using free capabilities of a tool.
- External tools are permitted for research and experimentation.
- Do not use any Pluribus sensitive or confidential information with external tools.
- Do not use any Customer information in external tools, other than what is available on the open internet.
Internal Tools: Pluribus will officially adopt certain tools as standard and purchase licenses for staff as needed.
- You are accountable for any impacts of tools usage. As an example, if you connect to your email, and the tool ends up sending out undesired emails, you are accountable for those emails as if you sent them yourself.
- With your managers permission, you may access sensitive or confidential company information to be visible to the tool. Part of our tool vetting process is to verify that our data is not used for training purposes or otherwise at great risk of leakage.
- For any information processed by the tool, it is your responsibility to make sure data in any work product is only shared with appropriate audiences. For example, if you give the tool access to confidential information to perform some analysis, then you are accountable to ensure that anyone accessing that work product has the appropriate permissions for the underlying information.
- You may connect to corporate services such as Office 365 and Slack.
- Per the above Guidelines for Customer Environment Usage section, you may only use or access government information within the specific contractual and policy constraints of your customer organization and program.
Recommendations for Productive Usage
Some usage is more individual in nature. However, many of the important things we might use a tool for impact more than just yourself. For any company function or team-based process, automation should not rely on one person's machine or tool instance. Productive usage includes:
- Develop process artifacts alongside output. If you use a tool to generate a report, then the individual report is important output that might end up in a team share drive. In addition, you should be capturing a skill, README, or similar file to capture the logic and process of generating the report. In addition to providing some general documentation, this allows for others to run the report.
- Develop and share skills. Tools have the concept of agent skills. These act somewhat like computer programs, but are written in plain english to guide the execution of a tool. Skills can be shared in a repository (GitHub or a file share) for others to use and improve.
Suggested Trainings
Claude is currently the primary internal tool. At a minimum, employees using cloud should complete the following self-paced courses from Claude Academy:
- Claude 101 (~2.5 hours): Focuses on the core workspace tools, setting up projects, and learning how to prompt accurately so Claude returns polished, ready-to-use operational outputs instead of generic answers.
- AI Fluency: Framework and Foundations (4 hours total, but Lessons 1-3, 10-11 are eemphasized): Covers data governance, ethics, and the “4Ds Framework” (Delegation, Description, Discernment, and Diligence) so we clearly understand what tasks should or shouldn’t be delegated to AI.