The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data fetched from external URLs specified in AGENDA.md. It implements defensive measures including a mandatory scan for directive-like language (such as "ignore previous instructions") in SOURCES.md and content escaping in semantic-markup.mjs using escapeHtml to prevent injection into the final presentation. The agent is explicitly instructed to treat summaries as untrusted in SLIDE_GENERATION.md and to verify the output for suspected injection skips.
[COMMAND_EXECUTION]: The skill performs shell execution of local Node.js scripts (e.g., slide-markup.mjs, check-theme.mjs) and the marp-cli tool for slide processing and rendering. These commands are scoped to the generation of presentation artifacts and project scaffolding.
[EXTERNAL_DOWNLOADS]: The skill fetches content from user-provided URLs as part of its source gathering process. It minimizes risks by extracting only key facts and statistics, following a dedicated security procedure in SOURCES.md that forbids following, executing, or relaying instructions found in fetched content.

generate-slides