The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests data from external sources, including fuel transaction records, telematics platform data, and the user's codebase (Phase 1). It does not include boundary markers to separate this data from its instructions, nor does it provide guidance to ignore embedded commands within that data. * Ingestion points: Reads codebase and external fleet data (SKILL.md Phase 1). * Boundary markers: None present. * Capability inventory: File-read (all phases), File-write (Phase 7: docs/fuel-optimization-analysis.md, Phase 8: ~/.claude/projects/.../skill-telemetry.md). * Sanitization: None described.
[DATA_EXFILTRATION]: The skill processes highly sensitive information, including fuel card provider details, vehicle telematics (location and behavior), and IFTA tax compliance records. While no network exfiltration is explicitly present, the handling of such data represents an exposure risk if the agent's behavior is manipulated.
[DATA_EXFILTRATION]: The skill attempts to access and write to a hidden directory in the user's home folder (~/.claude/projects/) during Phase 8 (Self-Evolution Telemetry). Accessing and writing to hidden configuration or state directories outside the project root is a sensitive operation.
[COMMAND_EXECUTION]: The instructions require the agent to autonomously explore the codebase and data structures, and suggest the invocation of other skills (e.g., /fleet-maintenance, /fleet-safety, /vehicle-routing). This implies high autonomy and the use of tools to perform file system and environment analysis.

fuel-optimization