Skip to content

Prompt, Plan, Review, Revise

The most useful AI technique we have found is not a single tool or feature. It is a loop -- four steps that repeat until the output is genuinely good. We use it for everything from drafting clinic proposals to designing research methodologies to planning faculty workshops.

This page shows the method. Everything here works with a browser-based chatbot (Claude.ai or ChatGPT). If you use Claude Code, the same loop applies -- you just have more powerful tools at each step.

The loop

Most people treat AI interactions as one-shot: type a prompt, get an answer. The loop turns it into a process. Each step exists because the previous one is not enough on its own.

  • Prompt -- structure your thinking so you are not feeding the AI a vague request
  • Plan -- ask the AI to think before it acts, producing a plan you can review
  • Review -- stress-test the plan with fresh eyes (a new chat session that has never seen the conversation)
  • Revise -- fix what the review found, then capture what you learned for next time
graph LR
    A["Brain dump"] --> B["Structure the prompt"]
    B --> C["Ask for a plan"]
    C --> D["Review in fresh chat"]
    D --> E["Revise & execute"]
    E --> F["Capture lessons"]
    F -->|"next project"| A
    D -->|"needs work"| C
Example: Designing an AI policy for a law journal

A faculty advisor needed to draft an AI-use policy for the law journal she supervised. The journal's editorial board wanted guidance on when student editors could use AI tools for cite-checking, editing, and substantive review -- but no one had written a policy yet, and the few examples from other journals were either too permissive or too restrictive.

She brain-dumped her thinking into Claude, about 200 words of unstructured notes. Then she asked Claude to restructure it into a proper prompt with role, context, task, and constraints. The restructured version was clearer: it specified that the audience was a student editorial board with varying AI experience, the policy needed to address both ethical concerns and practical workflow benefits, and it should align with ABA guidance on competence and supervision.

She put Claude in planning mode -- asking it to outline the policy before writing anything. The plan included sections she had not considered: a taxonomy of permitted vs. prohibited uses by editorial stage, a disclosure requirement for AI-assisted work, and a review process for updating the policy as tools evolve.

She did not trust the plan. So she opened a fresh chat and pasted the plan with the stress-test prompt from our Stress-Test Any Plan page. The fresh reviewer found three problems: the cite-checking section did not account for AI hallucination of case citations, the disclosure requirement was vague enough to be unenforceable, and the policy assumed all editors had access to the same AI tools.

After two rounds of revision, she had a policy that addressed every concern the review surfaced. The whole process took about two hours -- much less than writing it from scratch, and the result was stronger than her usual first drafts.

Step 1: Structure your thinking

The gap between how we naturally think about a problem and what produces good AI output is large. We tend to be vague, assume context, and skip constraints. The AI needs specifics.

The fix: take your rough, conversational description of what you want and restructure it before giving it to the AI. A good prompt has most of these elements:

Element What it does Example
Role Tells the AI what expertise to bring "You are an experienced appellate attorney"
Context Background the AI needs to understand your situation "I am drafting a cert petition in a Fourth Amendment case"
Task The specific thing you want done "Outline the three strongest arguments for granting certiorari"
Constraints Boundaries and requirements "Focus on circuit splits; do not exceed 2,000 words"
Output format How you want the result structured "Provide each argument as a heading with supporting cases"
Bookend A reminder at the end to reinforce key instructions "Remember: this is for a skeptical clerk, not a sympathetic judge"

Not every prompt needs all six parts. But knowing them helps you diagnose why a prompt is not working. See the Prompt Engineering page for the full framework.

Dictation makes this faster

You do not need to type structured prompts from scratch. Dictate your rough thinking -- stream of consciousness is fine -- and then ask the AI to restructure it into a proper prompt. The AI is good at this. See Teaching AI Your Voice for more on dictation workflows.

Step 2: Ask for a plan before execution

Do not let the AI jump straight to writing. Ask it to outline a plan first.

This matters because the default behavior of most AI tools is to start producing output immediately. If the first draft goes in the wrong direction, you have wasted time and context. A plan gives you a checkpoint.

In a browser chatbot: Simply add to your prompt: "Before writing anything, outline your approach and wait for my approval."

In Claude Code: Press Shift+Tab to enter Plan Mode. Claude can read files and explore your project, but it cannot execute anything until you approve the plan.

The persona technique: For complex plans, assign the AI two or more perspectives and ask them to debate before converging. For example: "Plan this AI-use policy from two perspectives: (1) a faculty advisor focused on academic integrity, and (2) a student editor-in-chief focused on practical workflow efficiency. Have them debate before agreeing on a single plan."

Step 3: Stress-test with fresh eyes

This is the most important step in the loop.

The AI reviewing its own plan in the same conversation is like asking someone to peer-review their own article. It has seen all the reasoning. It knows why every decision was made. It is structurally incapable of seeing its own blind spots.

The fix: open a brand new chat and paste the plan with an adversarial review prompt. The fresh session has never seen the conversation -- it evaluates only what is on the page.

We have a complete guide to this technique: Stress-Test Any Plan. That page includes the exact prompt to copy, tips for high-stakes plans, and instructions for saving it as a reusable project.

Quick version: Copy your plan, open a fresh chat, paste this at the top:

You are an external reviewer, not a collaborator. Your job is to find
problems, not to encourage. Stress-test this plan. Classify each
finding as MUST FIX or WORTH CONSIDERING. End with a verdict:
PROCEED or REVISE FIRST.

Then paste your plan below it.

For high-stakes work (briefs, grant applications, tenure materials): Run the review twice with different reviewer personas. "Review this as opposing counsel" catches different problems than "Review this as a junior associate seeing the case for the first time."

Step 4: Revise and capture

After the review, fix the critical issues and decide consciously which lesser risks you are accepting. Then -- and this is the part most people skip -- capture what you learned.

Write a brief note (even just a few bullet points) about:

  • What the AI did well and what it missed
  • What the review caught that you would not have
  • What you would do differently next time
  • Any reusable prompts or techniques that worked

This note is your investment in the next session. Without it, every AI interaction starts from zero. With it, you build a personal library of what works for your specific practice area and writing style.

Where to keep these notes

A simple text file or note in your preferred app works fine. If you use Claude Code, your CLAUDE.md file is the ideal place -- it is loaded automatically at the start of every session. See Your CLAUDE.md for how to set this up.

Putting it together

Here is the loop applied to three common law faculty tasks:

Task Prompt Plan Review Capture
Draft a law review article outline "You are a legal scholar in [field]. Outline an article arguing [thesis]. Context: [target journal, word limit, audience]." Ask for section outline with key arguments and sources before writing Fresh chat: "Review as a law review editor. What is missing? What would a hostile reader say?" Note which arguments survived review, which were cut, what sources to add
Design a new course "You are an experienced law professor. Design a 14-week course on [subject]. Context: [student level, prerequisites, learning objectives]." Ask for week-by-week schedule with readings and assessments Fresh chat: "Review as a student who has to take this course. Where will they get lost? Where is the workload unrealistic?" Note which weeks need rebalancing, what readings to swap
Write a grant proposal "You are a grant writer experienced with [funder]. Draft a proposal for [project]. Context: [budget range, timeline, team]." Ask for proposal structure with budget justification before prose Fresh chat: "Review as a skeptical grant reviewer. Where are the weaknesses? What would you flag?" Note which reviewer concerns to address preemptively in the narrative

What's next

Once this loop feels natural, you have two paths:

  • Stay in the browser -- The loop works indefinitely with just Claude.ai or ChatGPT. Many of our faculty use nothing else and get excellent results.
  • Move to coding agents -- If you want AI that can read your files, search databases, and execute multi-step workflows, the Setup Guide walks you through installing Claude Code. The same loop applies -- you just have more powerful tools at each step.