Task List Generation for Observability
Summary
A technique for decomposing a PRD into a detailed, nested task list that serves three purposes beyond simple project planning: (1) observability — you can see exactly what the AI agent plans to do before it does it, (2) control — you approve the plan before execution, creating a human-in-the-loop gate between planning and action, and (3) stakeholder engagement surface — the task list is a readable artifact that technical partners, reviewers, or collaborators can inspect, critique, and contribute to without needing to understand the implementation.
The mechanism: a rule file instructs the AI to first propose a high-level plan and wait for approval, then generate a full task list with nested parent tasks, sub-tasks, and sub-sub-tasks as Markdown checkboxes. This makes the AI’s intended approach legible and editable before any code is written.
How to Apply
When to use: When you need visibility into how an AI agent will approach a complex task, or when you need to give others (collaborators, reviewers, PMs) a way to engage with the plan. Especially valuable when the AI is about to do multi-step work where early mistakes compound.
When not to use: Single-step or clearly-defined tasks where the execution path is obvious.
Steps:
- Create a
generate_tasks.mdrule file that defines how to decompose a PRD into a nested task list - @-include both the rule file and the PRD in a new chat
- The AI proposes a high-level plan first and asks for approval — this is the human-in-the-loop gate
- Once approved, the AI generates
TASKS.mdwith nested checkboxes: parent tasks → sub-tasks → sub-sub-tasks - Review the task list for correctness, ordering, and completeness before starting execution
Why this matters beyond planning: The task list is a communication artifact. A technical partner can read it and say “you’re missing a migration step” or “swap the order of steps 3 and 4” without needing to pair-program. This increases the surface area for collaboration — people can contribute to AI-assisted work without being the one driving the agent.
For AI PMs: This technique maps directly to how AI agent products should expose their planning. When an AI agent is about to take multi-step action on behalf of a user, showing the plan first (and waiting for approval) is a UX pattern worth studying. The Markdown checkbox format is also notable — it’s a lowest-common-denominator format that’s readable by humans, parseable by AI, and diffable in version control.
Sources
From: Ryan Carson’s Structured AI Development
Key quote: (on the AI proposing a high-level plan first and asking for approval) “This human-in-the-loop step makes sure the plan is headed in the right direction.”
Attribution: Ryan Carson
What this source adds: Carson’s generate_tasks.md rule file is the concrete implementation. The key design choice is the two-phase generation: high-level plan → approval → detailed tasks. This prevents the AI from going deep on a wrong approach. The Markdown checkbox format is simple but powerful — it creates a living document that tracks progress as the AI works.
Links: Original | Archive