Notes On Work - by Caleb Porzio AI workflow tool brainstorm session
Feb 27, 2026
A brainstorming session about building an AI-driven workflow tool that enforces doing one thing at a time. They outline steps like reproducing bugs, writing failing tests, and documenting problems without solutions. The process emphasizes iterative question loops, stripping solution language from findings, and preparing low-effort versus high-effort fixes. Practical execution details and parallelization ideas are explored.
AI Snips
Chapters
Transcript
Generate One Solution Per Loop Not Many At Once
- Prefer iterating single solutions in a loop rather than asking an LLM to enumerate many solutions at once.
- Caleb found listing many solutions at once broke OTAT; better to generate one solution per run and loop until coverage is sufficient.
Use Questions To Turn Unknowns Into Empirical Tasks
- After listing potential solutions, have the AI produce targeted questions whose answers would make the decision trivial.
- Caleb asks Claude to write questions (not answers) and then runs a looped agent to answer each question empirically.
Answer Questions One By One With Empirical Runs
- Run answers in a loop: pick one question, answer it empirically (run code/tests), write the result, then restart.
- Caleb implemented a loop where each AI session tackles a single question and appends its findings to questions.md.