Plan, implement, review

This is my current AI-assisted software engineering workflow. My aim with this workflow is to operate at a higher level of abstraction while staying close to the code.

Plan

1. Start an agent. Explain the problem to be solved and propose a solution based on my understanding of the project and software engineering principles. I use dictation to speak this into the computer because I find it easier to brain dump in this format.
2. Ask the agent to write a plan for how to implement this code change, asking clarifying questions if needed.
3. Review the plan and iterate on it together with the agent.
4. Once I'm satisfied with the plan, put the planning agent to one side.

Implement

1. Start a new agent and ask it to implement the plan. Starting a new agent here clears the context window, so more context is available to handle the implementation. This also allows me to use a different model for planning and implementation.
2. Wait for the agent to complete or stop with questions. I avoid monitoring the agent as it's working because I don't want to intervene. Instead, I'll review the output once it's finished.
3. If the agent has questions or gets stuck, re-consider and revise the plan. Clean the working tree and restart the agent using the revised plan. Restarting the implementation, rather than trying to fix it, helps me learn how to write plans that an agent can implement without intervention.

Review

When the agent completes, review the code. Does it work? Does it solve the problem? Is it a good implementation?

Then:

• If the implementation highlights a problem with the plan, discard the implementation and revise the plan.
• If there are minor issues with the code change, prompt the implementation agent to make specific changes, or make them myself.

Experiments

I'm trying out git worktrees as a way to run more than one instance of this workflow in parallel. I think this could allow me to stay in the planning loop while spinning off implementation agents.

I'm also experimenting with capturing common patterns from the review stage into a code review prompt that I could run before reviewing the code manually.

I'm testing ways for agents to let me know when they've completed a task or got stuck by using a CLI command to show a desktop notification.