Methods for Breaking Work Down for AI Agents

When working with agentic AI, providing a massive, monolithic task is a recipe for hallucinations and poor alignment. What you ultimately need is a structured to-do list of some sort. The AI needs to know not just what to do, but how the work is segmented.

Tools for Task Management

The Simple Approach: TODO.md

A plain text file (like TODO.md or a plan.md from the Continuous Alignment process) is often the most effective method. To get the most out of your TODO.md when working with AI agents, you should establish clear conventions for how tasks and subtasks are represented and tracked.

For example, you can define a system where:

• Top-level tasks are identified by a task ID (like T001), followed by a short description.
• Sub-tasks are listed underneath, starting with a dash (-).
• When the AI completes a sub-task, it replaces the dash with a pipe (|) to mark it as done.
• Top-level tasks are separated by exactly one blank line.

Here is an example of what a TODO.md file might look like for building a simple to-do web application using these rules:

T001: Initialize Project Structure
| Scaffold Next.js application
| Install Tailwind CSS
- Set up ESLint and Prettier

T002: Create Database Models
- Define Todo schema in Prisma
- Create database migration
- Write seed script for initial test data

T003: Build Frontend Components
- Create TodoItem component
- Create TodoList container
- Add "Create Todo" input form

By providing these rules in a conventions file (like an AGENTS.md file), you can instruct the AI to always make a plan with detailed subtasks, branch off for each task, commit incrementally after each subtask, and mark them as complete as it goes. This turns a simple text file into a powerful state-tracking mechanism.

External Issue Trackers

If you have access to them (and the AI has integration capabilities), using Jira or GitHub Issues can also work well.

The Core Requirement: A Project Plan

Regardless of the tool you use, the fundamental requirement is a project plan. You need a clear idea of the order in which things need to be implemented. Providing this sequenced roadmap allows the AI to focus on one discrete, verifiable step at a time, checking its work and aligning with your goals continuously.

Conclusion

Breaking down work isn't just about making tasks smaller; it's about creating a verifiable trail of progress. By structuring your requests into discrete tasks—whether through a TODO.md file or an issue tracker—you provide the AI with the boundaries and focus it needs to succeed. This reduces errors, minimizes rework, and keeps the AI closely aligned with your vision from start to finish.