The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on the obsidian CLI tool, which provides an eval command (obsidian eval code="..."). This capability allows the execution of arbitrary JavaScript code within the Obsidian application, posing a risk if the input is manipulated.
[PROMPT_INJECTION]: The skill features a "Teaching Loop" and "Auto-Save Rule" that automatically append new instructions to the core configuration file (claude.md) based on user corrections. This mechanism is vulnerable to indirect prompt injection from untrusted data sources.
Ingestion points: Web content via defuddle, meeting transcripts (Fireflies/manual), and vault-resident project files (SKILL.md).
Boundary markers: Absent; the skill does not implement delimiters or safety instructions to distinguish between processed data and agent instructions.
Capability inventory: File system read/write via the obsidian CLI and direct access, arbitrary JS execution via obsidian eval, and network fetching via defuddle (SKILL.md, references/obsidian-cli.md).
Sanitization: No sanitization or validation logic is defined for the data extracted from external sources or for the rules generated during the "Teaching Loop."
[EXTERNAL_DOWNLOADS]: The skill utilizes the defuddle CLI tool to fetch and parse content from external URLs. While intended for markdown extraction, this enables the agent to interact with and ingest data from any remote domain provided in the conversation (SKILL.md).

assistant