The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions include specific shell command templates that incorporate user-provided content. Specifically, in references/idea-gate.md, the agent is instructed to perform a local overlap check using grep -ril <key-terms> writing/ originals/ ideas/ 2>/dev/null. Since the <key-terms> are extracted from user-supplied idea sources (such as notes, raw text, or external URLs), this creates a risk of command injection if the agent or platform executes the shell command without proper escaping or sanitization of the input terms.
[PROMPT_INJECTION]: The skill features multiple ingestion points for untrusted data, establishing a surface for indirect prompt injection. The /write ref and /write distill modes in SKILL.md and the /write idea-gate mode in references/idea-gate.md extract content from user-provided URLs or raw text. The instructions lack explicit boundary markers or directives to ignore embedded instructions within the processed content. Combined with the agent's capabilities for network access and shell execution, this increases the risk of the agent being manipulated by malicious instructions hidden in the source material.
[EXTERNAL_DOWNLOADS]: The skill requires the ability to fetch and analyze content from arbitrary external URLs to support its 'Reference', 'Distill', and 'Idea Gate' functionalities. This introduces a dependency on external, untrusted network resources.
[NO_CODE]: The skill consists entirely of instructional Markdown and supporting reference files. No actual executable source code, such as Python or JavaScript scripts, was provided in the skill package for code-level static analysis.

write