Skip to content

Prompt, Plan, Review, Revise

Author/Source: Chris Blattman, claudeblattman.com

Key Ideas

  • The core workflow is a four-step loop: /prompt (structure your thinking), Plan mode (think before acting), /review-plan (stress-test with a fresh agent), and /done (capture session context for continuity).
  • Most people treat AI as one-shot interactions; the loop turns it into a repeatable process where each step compensates for the limitations of the previous one.
  • /prompt takes rough, stream-of-consciousness input and reformats it into a structured prompt with Role, Context, Task, Constraints, Output Format, and Bookend sections -- then executes it.
  • Plan mode (Shift+Tab in Claude Code) forces the AI to read, explore, and propose approaches without executing anything, creating a deliberate pause for review.
  • /review-plan spins up a separate agent in a fresh context that has never seen the conversation, providing genuinely independent critique -- like external peer review rather than self-review.
  • /done captures decisions, open questions, and follow-ups so the next session starts with context rather than from scratch, building a searchable history across sessions.
  • The persona technique assigns multiple expert perspectives (e.g., web designer + skills engineer) to debate and converge on a plan before committing.

Summary

This article describes a four-step iterative workflow for getting high-quality output from Claude Code. The method was demonstrated by building the claudeblattman.com website in 90 minutes on a plane. The author brain-dumped a rough description, used /prompt to structure it, assigned two personas (web designer and Claude Code skills engineer) to debate the plan in Plan mode, then ran /review-plan twice with different reviewer personas to catch blind spots.

The intellectual centerpiece is the /review-plan step, which addresses a fundamental limitation of AI self-review: an AI critiquing its own plan in the same conversation has seen all the reasoning and cannot identify its own blind spots. By spinning up a fresh agent that reads the plan cold, the workflow mimics external peer review. The article provides concrete examples of what these reviews catch -- missing prerequisites, assumed knowledge, structural errors, and usability problems. Each review round takes about five minutes.

The /done step addresses session continuity. It writes structured handoff notes so future sessions begin with context. The project variant anchors the handoff to a specific directory and auto-loads when Claude Code opens in that directory. Combined with a CLAUDE.md file that provides persistent personal context, the loop creates a system where AI interactions compound rather than starting fresh each time.

Relevance to Economics Research

The prompt-plan-review-revise loop maps naturally onto research workflows. Economists can use /prompt to structure rough ideas for grant proposals or paper sections, Plan mode to outline an empirical strategy before committing to code, /review-plan to stress-test a research design with simulated referee personas, and /done to maintain continuity across multi-week projects. The persona technique is especially relevant: assigning "methodologist" and "applied researcher" personas to debate an identification strategy before implementation mirrors the value of co-author feedback in the research process.