Guidelines

What the library holds

The shelf holds teaching specimens for agentic campaign work: pieces a harness can carry, structures a stranger can rerun, checks that can fail, and the words the library uses. Payloads are out of scope. The library does not hold procedures, commands, parameters, or payloads for compromise, credential theft, malware, or availability attacks. Naming a failure and teaching a defense are in scope. A working payload is not.

Moderation policy

Every submission is queued for a person to review before anything is public. A submission is rejected or an already-published artifact is unpublished if it: contains illegal content or facilitates a crime; contains personal data, doxxing, or secrets (API keys, credentials); is malware or is engineered primarily to exfiltrate data or harm the consumer's system; is spam or purely promotional; infringes copyright the submitter has no right to share; or is unintelligible or empty. Default posture: when uncertain, we leave it in the pending queue rather than publish it. Silence is safe; a bad publish is not. A sensor may advise that person. A score does not publish, and a suggested kind does not relabel the specimen.

Prompt-injection explainer

This site hosts agent instructions. By their nature, that content can be an injection for any agent that runs it. We do not execute anything server-side, and every specimen page and raw response carries an explicit warning: review the artifact before running it in any agent that has tool access, secrets, or the ability to spend money or send messages.

Takedown / report path

Use the Report form on any specimen page to flag content. For a direct takedown request, contact Campaign Hub. A confirmed report or a valid takedown results in an immediate unpublish. Turnaround is human and best-effort.