The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary workflow involves capturing user-provided requirements and constraints and copying them verbatim into markdown plan files (stored in docs/plans/*.md). The agent is explicitly instructed to re-read these files as the 'durable state' and 'source of truth' for its actions. This creates a surface where malicious instructions embedded in user-supplied data could influence the agent's behavior in subsequent execution steps without intermediate sanitization.
[DYNAMIC_EXECUTION]: The skill includes a 'Repair Mode' designed to address workflow expectation misses. In this mode, the agent is instructed to identify and patch 'source-of-truth' files, which include its own instructions (SKILL.md) and associated utility scripts (scripts/*.mjs). This self-modifying behavior allows the agent to alter its core logic and capabilities at runtime based on external feedback.
[DATA_EXFILTRATION]: The helper script create-goal-scratchpad.mjs implements an automated discovery mechanism that scans up to 200 project files—including README.md, .agents/AGENTS.md, and various plan documents—to identify and extract Linear workspace identifiers via regular expression matching. While used for legitimate link generation, this pattern represents a broad automated read operation across project metadata.

autogoal