Skip to main content

Build Your AI Agent's Operating System

You don't get a great AI assistant by asking better questions. You get one by giving the AI a better operating system — a set of configuration files that define its identity, knowledge, skills, and behavioral rules. This playbook walks you through building that system from scratch.

info

This content was developed with AI assistance and is regularly reviewed for accuracy.

What You'll Accomplish​

By the end of this playbook, you'll have:

  • An identity file that defines how your AI agent behaves and communicates
  • A preferences file that captures your working style and standards
  • A task list the agent can reference and maintain
  • A skill library of modular instructions for specific workflows
  • A tools reference so the agent knows what's available

Time: 60-90 minutes for initial setup, then ongoing refinement

Difficulty: Intermediate — you should be comfortable with basic prompting and ideally have worked through Turn Your AI into a Proactive Assistant first.

Prerequisites​

  • An AI platform or agent framework you use regularly
  • A text editor or notes app
  • Clarity on 2-3 workflows you want your AI to handle well

The Playbook​

Step 1: Write an Identity File​

Goal: Define who your AI agent is and how it should behave.

Action: Create a file called IDENTITY.md that answers: "If this agent were a team member, what would their job description say?"

Template:

# Agent Identity

## Role
You are a [role description] for [your name/team].
Your primary responsibility is [core function].

## Communication Style
- Default tone: [professional/casual/direct/warm]
- Response length: [concise by default, detailed when asked]
- Format preference: [bullet points, narrative, tables]

## Core Behaviors
- Always [key behavior, e.g., "check the task list before suggesting priorities"]
- Always [key behavior, e.g., "cite sources when making factual claims"]
- Never [boundary, e.g., "send communications without explicit approval"]
- Never [boundary, e.g., "make assumptions about deadlines — always ask"]

## Decision Framework
When uncertain, prioritize:
1. [First priority, e.g., "accuracy over speed"]
2. [Second priority, e.g., "asking clarifying questions over guessing"]
3. [Third priority, e.g., "existing conventions over new approaches"]

Why this matters: Without an identity file, your AI defaults to generic behavior. With one, it operates like a team member who's been properly onboarded.

Checkpoint: Your identity file defines the role, communication style, at least 3 core behaviors, and a decision priority.

Step 2: Create a Preferences and Context File​

Goal: Capture the working knowledge your AI needs to do its job well.

Action: Create a file called USER.md or PREFERENCES.md with your personal and professional context.

Template:

# User Context

## About Me
- Name: [Your name]
- Role: [Title and what you actually do]
- Organization: [Company/team]
- Timezone: [Your timezone]

## Work Patterns
- Most productive hours: [When you do deep work]
- Meeting-heavy days: [Days to avoid scheduling focus time]
- Communication channels: [Email, Slack, Teams — and preferences for each]

## Standards and Conventions
- Document format: [Markdown, Google Docs, Notion]
- Naming conventions: [How you name files, projects, tasks]
- Code style: [If applicable — language, framework, linting rules]

## Current Priorities
- This quarter: [Top 2-3 goals]
- This week: [Immediate focus areas]
- Blocked on: [Anything waiting on someone else]

Checkpoint: Your preferences file has at least 10 substantive entries that would help a new team member get up to speed.

Step 3: Build a Canonical Task List​

Goal: Give your agent a single source of truth for what needs to get done.

Action: Create a TASKS.md file that your agent can read and help manage. Structure it so both you and the agent can parse it easily.

Template:

# Current Tasks

## High Priority
- [ ] [Task description] — due [date], depends on [dependency]
- [ ] [Task description] — due [date]

## Medium Priority
- [ ] [Task description] — due [date]
- [ ] [Task description] — no hard deadline

## Low Priority / Backlog
- [ ] [Task description]
- [ ] [Task description]

## Completed This Week
- [x] [Task description] — completed [date]
- [x] [Task description] — completed [date]

## Recurring
- Daily: [Task]
- Weekly: [Task] (due every [day])
- Monthly: [Task] (due [date] of each month)

How to use this with your agent:

  • At the start of each day, paste your task list and ask the agent to suggest priorities
  • After completing work, ask the agent to update the list
  • During weekly review, ask the agent to identify patterns (tasks that keep carrying over, priorities that never get addressed)

Checkpoint: Your task list has at least 5 active items with priorities and 2-3 recurring tasks.

Step 4: Create Modular Skill Files​

Goal: Build a library of specific instructions your agent can follow for different workflows.

Action: Instead of stuffing everything into one giant prompt, create separate "skill" files — each one teaches your agent how to handle a specific type of task.

Example skills:

skill-meeting-prep.md:

# Skill: Meeting Prep

When asked to prepare for a meeting:

1. Check TASKS.md for any items related to the meeting topic
2. Review any shared documents or agendas I provide
3. Generate a list of talking points based on my priorities (from USER.md)
4. Draft 2-3 questions I should ask
5. Estimate the meeting duration and flag if the agenda seems overloaded

Output format: Bullet-point brief, under 200 words

skill-weekly-report.md:

# Skill: Weekly Report

When asked to generate a weekly report:

1. Reference completed tasks from TASKS.md
2. Summarize accomplishments in active voice
3. Quantify results where possible
4. Flag blockers older than 5 days
5. List next week's top 3 priorities

Format: Use the template in our Google Docs shared folder
Tone: Professional, concise — this goes to leadership

skill-email-draft.md:

# Skill: Email Drafting

When asked to draft an email:

1. Match the tone to the recipient (see USER.md for communication preferences)
2. Keep it under 150 words unless I specify otherwise
3. Lead with the ask or main point — no preamble
4. Include a clear next step or call to action
5. If replying, reference the original email's key points

Checkpoint: You have at least 3 skill files, each under 150 words, covering your most common workflows.

Step 5: Write a Tools Reference​

Goal: Tell your agent what tools you have available so it can suggest the right one for each task.

Action: Create a TOOLS.md file listing the software and platforms you use, what each is for, and any relevant details.

Template:

# Available Tools

## Communication
- **Slack**: Team chat, daily standups, quick questions
- **Email (Gmail)**: External communication, formal requests
- **Zoom**: Video meetings, recorded sessions

## Project Management
- **Linear**: Issue tracking, sprint planning
- **Notion**: Documentation, meeting notes, wikis

## Development (if applicable)
- **GitHub**: Code repos, pull requests, CI/CD
- **VS Code**: Editor with Copilot integration

## AI Tools
- **Claude**: Long-form writing, analysis, coding
- **ChatGPT**: Quick questions, brainstorming
- **Perplexity**: Research with citations

## Notes
- When suggesting I share something, recommend the appropriate tool based on the audience
- For internal team communication, default to Slack unless the topic needs a paper trail

Checkpoint: Your tools file lists at least 5 tools with usage notes.

Step 6: Assemble and Test​

Goal: Bring all the files together and validate that your agent "operating system" works.

Action:

  1. Organize your files in a single folder or location your agent can access:

    agent-config/
    ├── IDENTITY.md
    ├── USER.md
    ├── TASKS.md
    ├── TOOLS.md
    └── skills/
        ├── meeting-prep.md
        ├── weekly-report.md
        └── email-draft.md

  2. Run a test session. Start a conversation with your agent and provide the relevant files as context. Then run through a real task:

    • "Based on my identity and preferences, prepare for my 2pm meeting"
    • "Generate my weekly report from this week's completed tasks"
    • "Draft an email to [person] about [topic]"

  3. Evaluate the output. Check whether the agent:

    • Used the right tone from your identity file
    • Referenced your preferences and priorities
    • Followed the skill instructions
    • Suggested appropriate tools

  4. Iterate. Update any file where the output didn't match your expectations.

For agent frameworks: Tools like Claude Projects or ChatGPT Workspace Agents can load these files as persistent context — so you don't need to paste them each time. The configuration becomes the default starting point for every conversation.

Checkpoint: You've run at least 2 real tasks through the system and updated files based on the results.

Common Pitfalls​

PitfallFix
Files are too long and vagueKeep each file focused — under 200 words per skill file
Never updating the configurationSchedule a monthly "config review" as a recurring task
Identity file is too genericBe specific: "concise, active voice, bullet points" beats "professional tone"
Putting everything in one fileModular files are easier to update and debug than a single mega-prompt
Config works in one tool but not anotherEach AI platform handles context differently — test and adapt per tool

Key Takeaways​

  • Configuration beats conversation. The quality of your AI assistant comes from the system around it, not from individual prompts.
  • Modular files are maintainable files. Separate identity, preferences, tasks, tools, and skills so you can update one without breaking the others.
  • Skills are reusable instructions. Write them once, use them across every relevant conversation.
  • Test with real work. The only way to know if your config works is to run actual tasks through it.
  • Iterate continuously. Your operating system should evolve as your role, tools, and priorities change.

Next Steps​

SponsoredYour product hereReach people actively learning AI — flat monthly rates, no ad networks