AI systemsClaude

Two prompts to make a fresh Claude chat review your work

No worksheet. No setup project. Run Prompt 1 in the creator chat, open a completely new Claude chat, then paste the packet with Prompt 2. The new reviewer sees the work and evidence without inheriting the creator chat’s preferred verdict. 🌱

You provideThe completed work in your current Claude chat

Claude doesPrompt 1 creates a clean evidence packet and Prompt 2 makes a new chat try to falsify it

You receivePass/fail by criterion, exact evidence, corrections and human decisions

one useful run

Use these in two separate Claude chats 👩‍💻

Run Prompt 1 in the chat that created the work. Copy its review packet, open a completely new Claude chat, paste the packet, then run Prompt 2. That separation is the point.

Prompt 1: run this in the original creator chat

copy prompt 1

Create a clean review packet for a completely new Claude chat. Include only: (1) the final artifact verbatim, (2) the acceptance criteria, (3) authoritative sources or source files, (4) objective tests already run and their exact results, and (5) constraints the reviewer must respect. Exclude hidden reasoning, self-justification, a proposed grade, a preferred verdict, and any claim that is not supported by the packet. Label missing evidence clearly. Return one copyable block.

Prompt 2: open a new Claude chat, paste the packet, then run this

copy prompt 2

You are a fresh evidence reviewer. Review only the packet above. Begin by trying to falsify the artifact rather than improve its wording. For every acceptance criterion, return PASS, FAIL, or NOT PROVEN; cite the exact evidence, file, source, calculation, or test result supporting the verdict; and state the smallest correction for every failure. Do not infer missing evidence and do not reward fluency or confidence. Then provide the corrected artifact where possible and list only the decisions that require a human. For high-stakes health, legal, financial, security, privacy, or irreversible decisions, stop at a human approval gate.

The quality gate. A good result performs the work and returns the finished artifact or decision. It should not give you another blank framework to complete.

Cleaner context is not guaranteed independence. A normal Claude Code subagent starts in a fresh isolated context, but it still receives a delegation task, usually inherits the model, and can load project-level instructions such as CLAUDE.md. A new top-level chat creates a cleaner boundary. For high-stakes work, add objective tests and use a different model or human when stronger independence matters.

Source boundary 🔗

Product-specific mechanics above are tied to current first-party Anthropic guidance. The prompt itself is Cindy's practical implementation, not an Anthropic product guarantee.

Keep human approval where consequences matter. Health, legal, financial, security, privacy and irreversible decisions still need a human. No prompt or Skill makes an agent infallible or injection-proof.

Follow for more. I share AI systems and Claude builds at @cindiezhu, and every guide lives on the free guides page 🌱