The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill dynamically generates TypeScript code (script.ts and script.test.ts) by synthesizing logic from the agent's recent conversation history and executes it.
Evidence: Step 8 invokes bun test script.test.ts or $B skill test against a temporary directory containing the dynamically generated files.
Evidence: Step 7 and Step 10 involve writing executable logic to the local file system and subsequently committing it to a permanent execution path.
[COMMAND_EXECUTION]: The skill extensively uses shell-level commands to interact with the environment, capture web data, and manage the local skill registry.
Evidence: Step 4 executes $B goto and $B html to capture snapshots of external websites.
Evidence: Step 11 runs $B skill run to verify the functionality of the synthesized script.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it uses previous conversation data (which may include content from scraped websites) as the source for code synthesis.
Ingestion points: Conversation history (up to 10 agent turns) and previous /scrape tool outputs (SKILL.md, Step 1 and Step 3).
Boundary markers: The skill attempts to locate specific intent lines but lacks robust delimiters to prevent embedded instructions from influencing the synthesis.
Capability inventory: The agent can write files (Step 7) and execute code via the Bun runtime (Step 8).
Sanitization: There is no evidence of sanitization or escaping of the conversation content before it is interpolated into the executable TypeScript template.

skillify