The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill clones a repository from GitHub and installs it as a Python package.
Evidence: Clones https://github.com/NousResearch/pokemon-agent and installs it in a virtual environment.
Note: The source repository is owned by the skill's authoring organization.
[COMMAND_EXECUTION]: The skill executes multiple shell commands to set up the environment, run the game server, and establish network tunnels.
Evidence: Commands include git clone, uv venv, pip install -e .[pyboy], and pokemon-agent serve.
[DATA_EXFILTRATION]: The skill initiates an outbound reverse SSH tunnel to a public service to expose a local web port.
Evidence: Executes ssh -R 80:localhost:9876 ssh://nokey@localhost.run to provide an external URL for the game dashboard.
Context: This is used for a well-known tunneling service (localhost.run) to allow the user to view the AI's gameplay in their browser.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from the game emulator which could potentially influence agent behavior.
Ingestion points: Game RAM states (GET /state) and game screenshots (GET /screenshot) in SKILL.md.
Boundary markers: None specified; instructions do not explicitly tell the agent to ignore instructions embedded in game text.
Capability inventory: The skill has file-write access (/tmp/pokemon.png), network access (SSH tunnel), and shell command execution capabilities across the startup and gameplay loop.
Sanitization: No explicit sanitization of RAM-extracted strings or OCR/Vision analysis results before they are used in decision-making.

pokemon-player