darlisted

# DAR — Deep Adversarial Reviews Iterative adversarial review: alternating Claude self-review and OpenAI GPT-5.6 cross-model review, with synthesis at the end of every round and after all rounds complete. ## Privacy Warning **OpenAI rounds upload the target file content to OpenAI's API.** Do not use this skill on files containing secrets, credentials, PII, or proprietary data you cannot share with a third-party API. The content is transmitted over HTTPS and processed by OpenAI under their data usage policy: https://openai.com/policies/api-data-usage-policies ## Target File The target file is the artifact to review — a plan document, specification, design doc, or code file. The skill reads it, runs the adversarial review, and writes findings back to the target file. ## Arguments - **`/DAR`** (no argument): 3 rounds — self → OpenAI → synthesis (classic 3-cycle). - **`/DAR:N`** where N is 1-9: N rounds of (self + OpenAI), each round followed by synthesis, then a final synthesis after all rounds. /DAR:1 = one self round + one OpenAI round + synthesis. /DAR:5 = 5 self + 5 OpenAI + synthesis. - **`/DAR:0`**: Unbounded — repeats self + OpenAI until zero CRITICAL, HIGH, or MEDIUM findings remain, then synthesis. Low findings are listed but do not prevent completion. Safety cap at 20 rounds — after 20 rounds, you MUST ask the user whether to continue (do not silently stop). If the user says yes, continue one round at a time, asking again each round. If the user says no, stop an