The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill features an indirect prompt injection surface as it ingests untrusted data from repository context (files, logs, issue reports) to architect goals.
Ingestion points: SKILL.md (Step 3) and references/goal-patterns.md (Bug/Research context) specify reading repository instructions, plans, tests, scripts, issues, and logs.
Boundary markers: The skill structures output into the /goal command format but relies on target-specific evaluators for final boundary enforcement rather than implementing its own sanitization logic for interpolated strings.
Capability inventory: The skill itself contains no executable code or subprocess calls; it is entirely instructional and produces text artifacts.
Sanitization: SKILL.md (Step 6) mandates a manual audit of drafted goals for 'false completion, unverified assumptions, impossible checks, hidden side effects, and target-specific evaluator blind spots'.
[SAFE]: The skill explicitly adheres to the 'Least Privilege' principle, stating that Goal mode does not expand tool permissions, authorize destructive actions, or replace project-level safety instructions.
[SAFE]: The skill provides guidance on 'Activation' that prevents unauthorized execution, instructing the agent to return ready-to-submit commands for user review rather than attempting to start remote sessions or CLI processes autonomously.
[SAFE]: References to external sources point to official documentation from trusted organizations such as OpenAI and Anthropic, or well-known community repositories.

goal-prompting