The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/deck.py executes shell commands to run the palette.py tool. It uses shlex.quote to properly sanitize all inputs, including user-provided requests and context, effectively mitigating command injection risks.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input and external files, passing them to a downstream model-powered process.
Ingestion points: Untrusted data enters the system through the plan and edit commands in scripts/deck.py via the --request, --context, and --source arguments.
Boundary markers: The SKILL.md file provides clear instructions to the agent to pass input 'as-is' and avoid pre-processing. The script relies on system-level command-line escaping rather than prompt-level delimiters to manage data boundaries.
Capability inventory: The skill has capabilities for reading and writing local files, executing subprocesses, and performing network requests (in remote mode).
Sanitization: The script employs shlex.quote to escape all arguments before they are passed to the shell, preventing the data from being interpreted as commands.
[EXTERNAL_DOWNLOADS]: When configured in remote mode via the $PALETTE_URL environment variable, the skill downloads .pptx decks and .png slide previews from the specified server. The script includes a check to verify the integrity of the downloaded file size.

palette