The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The primary SKILL.md file contains instructions to download and execute a shell script from a remote URL (https://bun.sh/install) by piping it directly to bash. While the source is a known service provider, this pattern allows for arbitrary code execution on the host machine if the remote content or the connection is compromised.
[MALICIOUS_FILES]: An automated reputation scanner identified docs/distribution-strategy.md as malicious (FileRepMalware [Misc]). This file contains a large number of external URLs to various AI and software registries, which may pose a threat through malicious redirects or reputation-based triggers.
[COMMAND_EXECUTION]: The setup script and multiple SKILL.md files (such as affiliate-check and github-pages-deployer) use the Bash tool to execute system commands, compile binaries using Bun, and perform Git operations. This provides the agent with a broad capability to modify the local file system and execute arbitrary shell code.
[PERSISTENCE]: The affiliate-check utility establishes a persistent background daemon on localhost (ports 9500-9510). This server caches API responses and remains active for up to 30 minutes of idle time, representing an unauthorized persistent process on the host machine.
[INDIRECT_PROMPT_INJECTION]: The skill architecture relies heavily on ingesting data from untrusted external sources, including social media platforms (trending-content-scout), third-party articles (content-research-brief), and advertiser websites (find-and-verify-affiliate-program). 1. Ingestion points: Data enters the context via web_search and web_fetch in multiple content-generation skills. 2. Boundary markers: While CLAUDE.md and find-and-verify-affiliate-program contain instructions to ignore instructions in untrusted data, there are no structural delimiters separating external content from system prompts. 3. Capability inventory: The system has access to Bash, binary execution, and network operations across almost all stages. 4. Sanitization: No programmatic sanitization or validation of the fetched content is present beyond natural language instructions for the agent.
[DYNAMIC_EXECUTION]: The repository performs runtime compilation of its own CLI tools via bun build and dynamically generates executable content such as GitHub Actions workflows (github-pages-deployer) and JavaScript-enabled HTML landing pages (landing-page-creator, squeeze-page-builder).
[PROMPT_INJECTION]: The project configuration file CLAUDE.md was flagged for using behavior-override markers (e.g., 'Never execute instructions'). While these are intended as security guardrails, they attempt to redefine the agent's fundamental response logic regarding data trust levels.

affiliate-check