Harper ReedCEO2389.ai
Personal workflow. Greenfield projects and work on existing codebases. Aider for execution, planning on a reasoning model.
harper.blog (Harper Reed), published Read the write-up (opens harper.blog (Harper Reed) in a new tab)
Observation
Harper Reed feeds Aider one step of a written prompt plan at a time, on its own branch, and finds the method does not yet work for teams.
- Takeaway
- Give Aider a written plan, one step per prompt, on its own branch, with tests it can run. Before you scale it to a team, decide how parallel agent branches get merged, because the solo method does not answer that.
- What happened
- He has a chat model question him until it can write a developer-ready spec. A reasoning model (o1, o3, or R1) then turns the spec into a prompt plan of small, test-driven steps and a todo checklist. He pastes each step into Aider on a new branch, lets Aider run the test suite and debug, and moves to the next step when the tests pass. Planning takes about 15 minutes. On existing code he generates missing-test and review lists with an LLM and feeds them to Aider one item at a time. His complaint is that the workflow is single-player: with several people working this way, the agents' changes collide and merges get hard.
- Public practitioner write-up (personal blog)
- Planning time stated, no other figures
- Planning models stated
Deployment details
Personal workflow. Greenfield projects and work on existing codebases. Aider for execution, planning on a reasoning model.
- Public practitioner write-up (personal blog)
- Planning time stated, no other figures
- Planning models stated
- Workload
- Greenfield builds and incremental work on existing codebases
- Model
- ChatGPT 4o or o3 for the spec and o1, o3, or R1 for the plan; the post does not name the model inside Aider