RPI

Splitting the work into three distinct phases before any code is written: research what already exists in the repository, plan the change in writing, and only then implement mechanically. It came out of HumanLayer, by Dexter Horthy, in 2024, and was picked up by tools such as Goose.

The separation exists to attack the failures an agent commits when handed unstructured work: inventing an API that does not exist, solving the wrong problem, overflowing the context window, and growing the scope on its own. Each phase has a gate, and the gate is where the human reviews.

Its own author retired the method in 2026. The reason is measurable: frontier models reliably follow between 150 and 200 instructions, and the planning prompt alone held 85 — adding the system prompt and the tool definitions, teams blew past the budget, and half the time the model skipped the alignment conversation entirely. The successors are CRISPY, with seven stages, and QRSPI, with five.