The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The researcher agent is instructed to use WebSearch and WebFetch tools to verify factual claims, formulas, and citations against authoritative external sources such as Wikipedia and official documentation. These operations are performed to populate the local facts.json ledger, ensuring that the AI does not fabricate information during lessons.
[COMMAND_EXECUTION]: The skill includes several local shell scripts (e.g., tests/run-all.sh) and a Node.js script (templates/lesson-gate.mjs) used for contract testing and deterministic quality gating of educational artifacts. These scripts are executed by the agent via the Bash tool to validate that the teaching content adheres to the 'never-vague' contract and other pedagogical rules.
[INDIRECT_PROMPT_INJECTION]: The skill's primary interaction loop involves ingesting natural language 'explain-back' responses from the user (the learner). This creates an attack surface where a user could provide adversarial input to influence the agent's behavior.
Ingestion points: Learner responses are ingested by the tutor agent in agents/tutor.md (Step 5 of the teach loop) for grading against the 6-type gap rubric.
Boundary markers: The instructions do not specify explicit delimiters (like triple backticks) or 'ignore' instructions for the specific learner input block in the prompt logic.
Capability inventory: The agents in this skill have access to Write, Edit, Bash, WebSearch, and WebFetch tools.
Sanitization: No explicit sanitization or filtering of user input is implemented in the provided instruction set, though the skill's focus on pedagogical grading acts as an inherent filter.

supertutor