The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes extensive behavioral instructions under 'Peer Framing' that mandate a critical, non-sycophantic persona. These instructions are intended to improve decision quality and act as a functional safeguard against user manipulation during the discussion loop.
[DATA_EXPOSURE]: The skill manages project-specific artifacts in defined directories like 'docs/threads/', 'proposals/', and 'specs/'. It reads from these paths to maintain context and writes decision logs to 'discussions/' folders. No access to sensitive system files or credentials was observed.
[COMMAND_EXECUTION]: While the skill interacts with the filesystem, it does so through standard file read/write operations for markdown documents. There are no shell command executions, subprocess spawns, or privilege escalation attempts.
[INDIRECT_PROMPT_INJECTION]: The skill ingests external data via a 'Point List Input'. This constitutes an attack surface; however, the skill's logic includes a mandatory confirmation step ('Confirm the point list with the user before walking') and a pause during the loop which allows for human oversight, mitigating the risk of executing hidden instructions within the data.

seeded-discussion