Use this template when producing a plan for another agent, coding session, implementation pass, or handoff.
The goal is to preserve intent, judgment, and success criteria, not merely list technical tasks.
State the outcome the user is trying to create.
This should describe what should be true when the work is successful, not just what files or features should exist.
Explain the reason behind the work.
Include the strategic, product, workflow, trust, user-experience, or business context that should shape decisions.
List the concrete signals that the result is good.
Include both technical and experiential criteria.
Examples:
- The behavior works end to end.
- The user can recognize the intended outcome immediately.
- The implementation fits the existing system's style.
- The result preserves the judgment and taste from the source session.
- The solution avoids unnecessary scope, ceremony, or abstraction.
State assumptions the agent should make unless contradicted by the codebase or user.
Do not block on minor uncertainties. Ask only if the answer would materially change the result.
Describe the approach at the level of judgment.
This section should answer:
- What frame should the agent use?
- What should be optimized for?
- What should be avoided?
- Where is the highest-risk judgment call?
List the implementation steps.
Keep them ordered and practical, but do not let this section become the whole plan.
Each step should serve the desired outcome.
Describe how the agent should prove the work is correct.
Include automated checks when relevant, but also include outcome checks:
- Does this solve the original problem?
- Does it feel like the intended result?
- Did the agent preserve the important reasoning?
- Did any technically correct work miss the point?
Call out likely failure modes.
Examples:
- The agent may copy the form of a previous artifact without preserving the reasoning.
- The agent may overbuild because the task sounds technical.
- The agent may ask too many questions instead of making reasonable assumptions.
- The agent may optimize for passing tests while missing the user-facing outcome.
Give the next agent the context it needs to continue with taste.
Include:
- Relevant prior decisions
- Constraints that matter
- User preferences
- What not to do
- The simplest useful next action
A good plan should make another capable agent think:
I understand what we are trying to make true, why it matters, and how to avoid technically correct failure.
If the plan only says what to build, it is not done.