The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to download an installation script from https://release.anza.xyz/stable/install. While Anza is a recognized entity in the Solana ecosystem, the source is external to the vendor.
[COMMAND_EXECUTION]: The skill uses a sh -c command with command substitution to execute the downloaded script directly in the shell. This pattern is commonly used for tool installation but represents a potential remote code execution vector.
[CREDENTIALS_UNSAFE]: The skill manages Solana private keys (keypair.json). It provides instructions on generating these keys and setting them as secrets. Although it includes best practices such as using .gitignore and temporary shell variables to avoid history logging, the handling of raw private keys carries inherent risk.
[PROMPT_INJECTION]: An indirect prompt injection surface is present in Step 1 (Configuration interview), where user-provided strings are captured and interpolated into configuration files and shell commands. * Ingestion points: User input for app name and program configuration (SKILL.md). * Boundary markers: None present. * Capability inventory: npm install, solana-keygen, goldsky compose deploy, git commit, gh repo create, and curl (SKILL.md). * Sanitization: No sanitization or validation of the user-provided configuration strings is specified.

compose-solana-transactions-setup