Skip to content

Stress-Test Any Plan with a Fresh AI Chat

No Claude Code required

Use Claude Code? You can do this from the terminal too

The workflow on this page works entirely in a browser -- Claude.ai, ChatGPT, or any AI chat tool. If you have Claude Code installed, you can run the same process from the terminal by pasting the prompt below directly into a fresh session. The key insight is the same either way: the reviewer must not be the author.

Why a fresh chat?

AI tools are great at helping you build plans. They are bad at critiquing them -- because the same model that helped develop an idea has a natural bias toward defending it. Researchers call this sycophancy: the tendency to agree with whatever the user seems to want (Anthropic, 2024).

The fix is simple: finish your plan in one conversation, then paste it into a brand new chat for critique. The fresh context eliminates anchoring bias -- the reviewer has zero knowledge of your reasoning, compromises, or earlier drafts. It is the AI equivalent of handing your draft to a colleague who was not in the room when you wrote it.

When this is worth doing

  • Brief strategies and motions -- before filing, get a structured adversarial review
  • Regulatory compliance plans -- surface gaps in your analysis before the agency does
  • Clinic proposals -- stress-test a new clinic structure before presenting it to the dean
  • Course redesigns -- catch logistical problems in a new curriculum before the semester starts
  • Grant applications -- find weaknesses before the review committee does
  • Hiring plans and committee proposals -- any faculty governance decision worth getting right

Skip for: routine emails, reversible decisions, plans you will iterate on daily anyway.

The method

1. Finish your plan. Get it to a state where you think it is ready -- in any tool, on paper, wherever.

2. Copy it with enough context. Include: the goal, deadline, available resources, key constraints, stakeholders involved, and what success looks like. A 3-sentence sketch produces a shallow review. Aim for 200+ words with concrete steps.

3. Open a new chat in Claude.ai, ChatGPT, or any AI tool. Do not continue in the conversation where you built the plan.

4. Paste the prompt below, then paste your plan at the bottom where indicated.

5. Read the critique, then iterate. Fix critical issues first. Ask follow-up questions. For high-stakes plans, open another fresh chat and run a second review from a different angle.

The prompt

Copy this into your fresh chat, then paste your plan at the bottom:

You are an external reviewer, not a collaborator. Your job is to find
problems, not to encourage.

Stress-test this plan. For each issue, point to the specific step or
sentence that triggered your concern. Classify each finding as:
- MUST FIX -- will likely cause failure or major rework if not addressed
- WORTH CONSIDERING -- creates risk, but the plan can proceed if you
  decide consciously to accept it

Find:
1. The top 3 ways this plan could fail -- work backward from failure
2. What's missing that someone experienced in this domain would expect
   (stakeholders, dependencies, resources, sequencing issues)
3. What assumptions or timelines look unrealistic -- and what evidence
   would confirm or refute them
4. What should change before proceeding

If domain-specific best practices matter, search for them and cite
what you find.

Do not rewrite the plan. Diagnose first.

End with a verdict: PROCEED or REVISE FIRST.

THE PLAN:
[paste here]

Use the strongest thinking model available

Heavier reasoning models produce sharper critique. Pick the most capable model available on your subscription -- the adversarial framing in this prompt does the rest.

What to do with the output

If the verdict is REVISE FIRST: Address the "Must Fix" items. Then ask: "Here is my revised plan addressing your concerns. Review it again."

If the verdict is PROCEED: Scan the "Worth Considering" items. Decide consciously which risks you are accepting and which are worth a quick fix.

For high-stakes plans (briefs in complex litigation, clinic accreditation proposals, tenure materials): Open a second fresh chat with a modified role -- "Review this plan as a [opposing counsel / accreditation reviewer / junior associate seeing this for the first time]." Different personas catch different blind spots. For maximum independence, use separate conversations rather than asking for multiple personas in one chat.

Common mistakes

Reviewing in the same conversation where you built the plan. The model will hedge its criticism because it participated in creation. Always use a fresh chat.

Trusting the verdict without reading the findings. The classification is a starting point. Read each finding and decide for yourself whether it applies to your situation. AI reviewers can be overconfident about risks that do not apply to your context (ScienceDirect, 2025).

Using a plan that is too vague. A 3-sentence sketch gets a 3-sentence review. Include concrete steps, timelines, resources, and constraints.

Adopting the AI's revision wholesale. When you ask for a revised plan, the model produces something plausible and polished. Resist adopting it without thinking. The revision is a starting point, not a replacement for your professional judgment (PMC, 2023).

Save this for repeat use

If you review plans regularly, save the prompt as a persistent project instead of pasting it each time.

Claude.ai Project

  1. Go to claude.ai/projects (requires Pro, Team, or Enterprise)
  2. Click "+ New Project"
  3. Name it "Plan Reviewer"
  4. Click "Set Custom Instructions" and paste the prompt from above
  5. Optionally upload reference documents (your firm's brief standards, past post-mortems, regulatory checklists)

To use: open the project, start a new conversation each time, paste your plan. The instructions apply automatically. Each new conversation gets fresh context -- no cross-contamination between reviews. (Claude Help Center -- Projects)

ChatGPT Custom GPT

  1. Go to chatgpt.com/create (requires Plus, Team, or Enterprise)
  2. Click the Configure tab
  3. Set name: Plan Reviewer; paste the prompt into Instructions
  4. Enable Web Browsing; disable DALL-E and Code Interpreter
  5. Save and choose "Anyone with a link" to share with colleagues

To use: start a new conversation with the GPT each time. Colleagues can use it immediately via the shared link -- no setup on their end. (OpenAI -- Creating a GPT)

Tips for both

  • Always start a new conversation for each plan -- reusing an old conversation reintroduces the anchoring problem
  • Keep a versioned copy of your instructions in a local text file with dates and change notes
  • Test with one plan you know is solid and one you know has problems before sharing with colleagues
Sources and references

Current as of April 2026. Model names and capabilities change frequently.