The Agent Skills Directory

[COMMAND_EXECUTION]: The skill interacts with the Binance Smart Chain by calling the baw (Binance Agentic Wallet) CLI tool. All state-changing operations follow a mandatory security workflow: a transaction preview is generated and simulated, risks are presented to the user, and a human-in-the-loop confirmation is required before the transaction is executed. This prevents the agent from performing unauthorized or unintended financial operations.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data from JSON files (binding.json and ticket.json). To mitigate risks of malicious data influencing agent behavior, the skill implements strict schema validation, including regex patterns for strings, length limits for metadata, and cryptographic signature verification for binding claims. Furthermore, the final output shown to the user is the raw transaction data and simulation results, which are not executed without manual approval.
[EXTERNAL_DOWNLOADS]: The skill relies on the viem library for blockchain communication and standard development tools. No suspicious or unverifiable external scripts or packages are downloaded at runtime.
[DATA_EXFILTRATION]: The skill uses environment variables for configuration (e.g., RPC_URL, AGENT_WALLET). These are used solely for their intended purpose of communicating with the Binance Smart Chain. There is no evidence of sensitive data being sent to unauthorized third-party servers.
[PROMPT_INJECTION]: The instructions in SKILL.md are descriptive and technical. They do not contain any instructions intended to override AI safety guidelines, bypass constraints, or extract system prompts.

realbond