The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user input and processes repository data (such as wiki pages and source code) to generate design specifications. This creates a surface for indirect prompt injection. However, the skill implements extensive mitigations, including a 'two-key execution handshake,' strict schema validation (validate-spec --strict), and a 'fork taste test' where a separate, fresh-context agent reviews the draft against the original user goal to detect drift or contradictions.\n- [COMMAND_EXECUTION]: The skill executes a bundled Python utility (scripts/mold.pyz) to perform internal logic such as artifact path resolution, task counting, and specification finalization. These commands are restricted to the skill's own local scripts and do not involve arbitrary shell execution of user-supplied strings.\n- [EXTERNAL_DOWNLOADS]: The skill can invoke external research tools to gather documentation or evidence for design grounding. These operations are part of the intended functionality and are explicitly logged in the 'Grounding' section of the generated specifications for user audit.

mold