The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources which creates a surface for indirect prompt injection attacks. If an adversary includes malicious instructions in team chat messages or repository files (like README.md or commit messages), the agent could potentially execute those instructions while generating the summary.
Ingestion points: Data is ingested via hackathon chats read (team chat and help request threads) and git clone (repository README, manifest files, and commit history).
Boundary markers: The instructions do not define clear delimiters or escaping mechanisms to separate untrusted external content from the agent's internal instructions during the summary generation phase.
Capability inventory: The agent has the ability to execute shell commands (npm, git, gh, hackathon CLI), write to the local filesystem (cloning repositories), and access network resources.
Sanitization: While the skill includes a safety instruction to avoid printing secrets found in repositories, it does not include specific sanitization or filtering to prevent instruction injection from the ingested text.
[EXTERNAL_DOWNLOADS]: The skill installs the @antasphere/hackathon package globally from the npm registry. This is a vendor-provided tool necessary for the skill's functionality. It also downloads team repositories from GitHub based on the event's organization configuration.
[COMMAND_EXECUTION]: The skill relies heavily on executing external CLI tools to perform its tasks, including npm for installation, git for repository management, gh (GitHub CLI) for invitation handling, and a custom hackathon CLI for platform interaction. The execution is scoped to the coach's authenticated context.

hackathon-coach-brief