The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection by interpolating untrusted user input directly into its instructions.
Ingestion points: User input is ingested via the $ARGUMENTS placeholder in SKILL.md (Step 2 and the final line).
Boundary markers: No delimiters or "ignore embedded instructions" warnings are used around the interpolated input.
Capability inventory: The skill has the capability to write to the filesystem using the obsidian_append_content tool.
Sanitization: No sanitization, escaping, or validation of the input is performed before interpolation.
[COMMAND_EXECUTION]: The skill utilizes dynamic context injection to execute a shell command when the skill is loaded.
Evidence: The !date command in SKILL.md executes at load time to substitute the current date into the document.
[COMMAND_EXECUTION]: The skill performs filesystem operations to read templates and write analysis results.
Evidence: The skill reads a template from ~/.config/opencode/templates/thinking-session.md.
Evidence: It writes content to a path determined by the $OBSIDIAN_PATH environment variable using the obsidian_append_content tool.

ultrathink