The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted text data that could potentially contain malicious instructions intended to influence the agent's analysis.
Ingestion points: The script scripts/build_match_context.py fetches match commentary from the FlashScore API and scrapes match previews from www.flashscore.co.
Boundary markers: The SKILL.md file uses clear headers and source tags (e.g., [API], [IND]) to organize data, but it lacks strict delimiters or instructions to ignore potential commands embedded within the external text.
Capability inventory: The skill possesses the capability to execute Python scripts, make network requests to arbitrary URLs, and write files to the local filesystem.
Sanitization: The script uses BeautifulSoup and regular expressions to clean HTML and pseudo-tags, which provides structural cleaning but does not prevent semantic prompt injection.
[COMMAND_EXECUTION]: The skill instructions explicitly direct the agent to execute bundled Python scripts (build_match_context.py and generate_teams_csv.py) with arguments derived from user input or discovery phases.
[DATA_EXPOSURE]: Although not a realization of exfiltration, the scripts are designed to read from sensitive local paths including ~/.claude/settings.json and ~/.claude/settings.local.json to retrieve the RAPIDAPI_KEY. While documented as a setup step for the Claude Code environment, automated access to configuration files containing other secrets is a sensitive pattern.

football-betting-analysis