The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The error-detector.sh script ingests the output of shell commands to identify errors and inject reminders into the agent's context. This represents a potential surface for indirect prompt injection if a command produces output designed to manipulate the agent's subsequent actions via the injected reminder.
Ingestion points: The scripts/error-detector.sh script reads tool execution results from standard input in JSON format (targeting tool_response or textResultForLlm fields).
Boundary markers: The resulting reminder is wrapped in <error-detected> tags when injected into the prompt context to help the model distinguish it from user input.
Capability inventory: The skill allows the agent to modify project documentation (CLAUDE.md, AGENTS.md), log entries to the .learnings/ directory, and create new skill scaffolds via the extract-skill.sh helper script.
Sanitization: The script utilizes standard system utilities (jq or Python's json module) for data extraction, which protects against basic payload-level injection during the parsing phase.

self-improvement