Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/xxyoudeadpunkxx/pa-pvp-mini/llms.txt

Use this file to discover all available pages before exploring further.

The two AI mode runs pa_pvp mini across two completely separate chat sessions — no shared memory, no connected context, no coordination between them. AI-A reviews the artifact and produces Round 1. AI-B receives the same artifact plus AI-A’s full output, independently reviews the material, and then audits the prior review. The result is a ping-pong: each pass applies fresh eyes to both the artifact itself and every judgment the previous AI made. Because AI-B has no prior commitment to AI-A’s conclusions, it is free to challenge them directly.

​
Why Two AIs

A single AI reviewing the same artifact twice will tend to anchor on its prior output. It has already formed conclusions, and re-reading the same material through the same model with the same context will often reproduce similar blind spots. A second AI — particularly one from a different model family — starts from scratch. It has no investment in the previous findings and no memory of the reasoning that produced them. This separation creates two independent views of the same artifact and one adversarial view of the prior review. The first AI finds structural problems in the work. The second AI can find structural problems in both the work and the first AI’s analysis. When both AIs independently flag the same issue, that convergence is strong evidence it deserves attention. When AI-B challenges a finding from AI-A, the disagreement itself is useful — it surfaces where the review logic was weak or the evidence was overstated.
Each chat operates completely independently. There is no shared memory between sessions. Every round carries the full prior context explicitly inside the PREVIOUS_ROUND_OUTPUT block, so no integration between chats is required — just copy and paste.

​
Step-by-Step Setup

1

Activate pa_pvp mini in AI-A

Paste the protocol text into your first AI chat session, or upload pa_pvp mini v2.3.txt if the chat supports file uploads. This sets the behavioral contract for the session.
2

Submit the artifact to AI-A

Send the material you want reviewed, labeled with the ORIGINAL_ARTIFACT block:
ORIGINAL_ARTIFACT:
[your material here]
AI-A runs Round 1 and returns a structured findings list with the ROUND header, FINDINGS block, and END OF PROTOCOL marker.
3

Open AI-B in a separate chat

Start a completely new chat session. This can be the same AI model or a different one. The only requirement is that it has no prior context from the AI-A session.
4

Activate pa_pvp mini in AI-B

Paste or upload the protocol into AI-B exactly as you did in AI-A. Each session must be activated independently.
5

Send the artifact and AI-A's output to AI-B

Submit both the original artifact and the full output from AI-A using labeled blocks:
ORIGINAL_ARTIFACT:
[same material as before]

PREVIOUS_ROUND_OUTPUT:
[paste AI-A's full output here, from ROUND: to END OF PROTOCOL]
AI-B detects that PREVIOUS_ROUND_OUTPUT is present and activates Round 2+ mode automatically.
6

AI-B runs Round 2

AI-B first reviews the artifact independently, then audits AI-A’s findings, and produces a merged fix status. Its output includes confirmed findings, challenged findings, new findings, and any explicit conflicts between its own view and AI-A’s.
7

Continue by alternating chats

To run additional rounds, copy AI-B’s full output and paste it into AI-A as the new PREVIOUS_ROUND_OUTPUT, alongside the same ORIGINAL_ARTIFACT. AI-A now runs Round 3. Repeat the cycle as many times as useful.

​
What AI-B Checks

AI-B checks two things, and the order matters. First: the artifact itself. AI-B reads ORIGINAL_ARTIFACT and runs a full independent review before looking at AI-A’s output. This is enforced by the protocol’s execution rules — the independent review phase must complete before the audit phase begins. This ensures AI-B’s own findings are not contaminated by prior framing. Second: the prior review. After forming its own view, AI-B examines AI-A’s findings and assesses each one against what it found independently. It confirms findings that hold up, challenges findings that do not, and surfaces problems AI-A missed. The label on each finding (INDEPENDENT, CONFIRMED, CHALLENGED, or NEW) records the outcome of this comparison directly in the output.

​
Starting from an Existing Review

You do not need a prior pa_pvp round to use two AI mode. If you already have a structured review — a list of findings, proposed fixes, tool output, notes from a colleague, or output from another review process — you can paste it as PREVIOUS_ROUND_OUTPUT. The AI will treat it as prior-round material: checking it against the artifact rather than accepting it. The PREVIOUS_ROUND_OUTPUT block accepts any structured review content. The reviewer will audit it under the same rules as any prior pa_pvp output: confirming what holds, challenging what does not, and identifying what it missed. 
ORIGINAL_ARTIFACT:
[your material]

PREVIOUS_ROUND_OUTPUT:
[existing review, tool output, or prior findings — not necessarily from pa_pvp]
Using two different AI model families — for example, one pass with GPT and one with Claude, or Claude and Gemini — maximizes the diversity of perspective. Different models have different training distributions, different tendencies toward certain categories of finding, and different thresholds for what they consider structurally significant. A finding that appears in both a GPT pass and a Claude pass is much harder to dismiss than one that appears in a single model’s output.

Build docs developers (and LLMs) love

Get started for freeTalk to us