The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill fetches content from external video platforms, including video titles, metadata, and top comments, to populate report.md and historical prediction files.
Ingestion points: Phase 3.5 fetches data from TikTok, Red (XHS), YouTube, Bilibili, and LinkedIn via platform-specific adapters.
Boundary markers: The instructions do not specify any delimiters or safety headers to prevent the agent from interpreting instructions embedded within the fetched comments or titles.
Capability inventory: The skill possesses Bash, Write, and Edit capabilities, allowing it to modify project files and execute local scripts based on the processed context.
Sanitization: No explicit sanitization or filtering of external content is described before interpolation into the project's markdown files.
[COMMAND_EXECUTION]: The skill utilizes the Bash tool to perform environment setup and validation.
Evidence: Phase 3 uses shell commands to copy templates and create directory structures. Phase 4 executes a test script (bash .cheat-hooks/session-start.sh) and uses command output to verify that hooks are correctly intercepting file edits.
[PERSISTENCE]: The setup process establishes persistent behavior modification by registering hooks in the agent's configuration.
Evidence: Phase 3, Step 6 modifies .claude/settings.json to register PreToolUse (for a "prediction lock") and SessionStart hooks. These hooks trigger shell scripts located in the project's .cheat-hooks/ directory whenever a new session starts or specific tools are invoked, ensuring the skill's logic persists across multiple agent interactions.

cheat-init