A solo developer documents a four-role harness of plan, work, critique, promote that commodified the frontier models for him
Scott Fryxell's post dated 2026-08-25 is a practitioner account of running three TUIs (Cursor, Claude, Pi) over one shared set of skills and a single AGENTS.md, which he says removed any dependence on a particular vendor's experience. His harness formalizes exploration into an explicit DAG task list, hands it to a worker that implements one node at a time, then brings in a separate critic to simplify and question the result, often pushing work back to the worker. He isolates each role after finding that a single prompt that plans, executes and critiques confuses its own objectives, and he built the pattern on Can Bölük's prewalk technique, which uses a frontier model for planning and the first task then hands off once the pattern is set. He runs deepseek-v4-flash for maintenance work and reaches for frontier models only on large refactors.
↳ Follow the thread