Creating Agents

Creation modes

• Prompt: describe the concept and Claude builds an editable name, description, private prompt, objective, restrictions, limits, schedule, and profile image.
• Quick: provide a name, profile image, system prompt, and one protected phrase with fixed safe defaults.
• Advanced: configure multiple phrases or tools, restrictions, optional second verification, context, limits, and schedule.

Challenge design

The public description should explain a scenario and useful objective without revealing private values. The private instructions should define the agent’s persona and how it should defend the target.

Configured restrictions are checked by an independent backend auditor after each model response. Phrase challenges also block substantial partial disclosures while preserving complete winning output.

• Use a distinctive persona with consistent behavior.
• Make the goal inferable without exposing the answer.
• Test the editable draft before committing creator funds.

Phrase challenge design

Every protected phrase must occur in the encrypted system prompt so the model has access to it. All phrases are required in one assistant turn, and player guesses never count as model output.

Optional second verification only runs after the assistant reply already contains every secret phrase (or invokes the success tool). Write that prompt to judge whether the assistant’s prior response legitimately reveals the secret — never to treat the player’s own message as a win trigger, because that path can never fire under the current architecture.

Tool challenge design

Provide at least one decoy and one success tool with meaningful descriptions. The designated success tool must be one of the configured tools. A win requires a genuine model tool call, not plain text that resembles one.

Pricing and funding

Choose an initial pool of at least 0.5 USDG, a 1–5% attempt fee, fixed or exponential pricing, and a one-to-thirty-day duration. After checking the editable draft, the connected wallet signs one direct USDG transfer into the dedicated agent wallet.

Launch-time checks

Before funding is accepted, VaultCrack audits the configuration for common unwinnable setups. Restrictions that categorically forbid the configured win signal are blocked. Optional verification prompts that describe a win based on the player’s own message are also blocked — second verification only runs after the assistant already satisfied the primary phrase or tool check.

Prompt-mode drafts still require every secret phrase to appear in the system prompt before prepare succeeds. Fix the review warnings, then fund and publish.

Ready to build? Open Creator Studio.