Skip to content

QRSPI — Question, Research, Spec, Plan, Implement

dexter-horthy's update to Research-Plan-Implement. He found that 10,000+ engineers had grabbed HumanLayer's RPI prompts but most were running them wrong — skipping the "work back and forth with me, starting with your open questions and outline" magic words, producing bad plans, then not catching the problems because they didn't read the code.

The update

Insert Question (an explicit outline-and-open-questions phase) and Spec (design discussion) before Plan:

  1. Q — Questions: model generates open questions and an outline; human resolves them.
  2. R — Research: facts-only extraction of how the code works today, no implementation opinions.
  3. S — Spec: design discussion — typically ~200 lines, high leverage for resteering.
  4. P — Plan: concrete steps, written only after S is approved.
  5. I — Implement: execute.

The two things RPI got wrong

  • "It's okay to not read the code." It is not. Read the code, not the 1000-line plan file.
  • Leverage was misapplied. People thought reading the plan saved time. It doesn't — if someone hands you a plan you didn't co-author, you have to re-read the code to verify it, defeating the point. Real leverage is in the design discussion where you can resteer cheaply.

The magic words trap

"Work back and forth with me starting with your open questions and outline before writing the plan."

Without that phrase, ~50% of users got unusable plans from Opus. Horthy's fix is to make the "Q" phase structural (a separate agent step) instead of prompt-dependent (a phrase users have to remember).

Why this matters

QRSPI is the distilled anti-pattern for 2026: "outsourcing thinking" is the failure mode, not slow engineers. Maps directly onto do-not-outsource-thinking and harness-engineering — the harness should structurally prevent the bad path, not rely on prompt hygiene.

Connects to