The Agent Skills Directory

[PROMPT_INJECTION]: The skill is subject to indirect prompt injection because it processes session history and logs that reflect user-generated content.
Ingestion points: Reads from agent-improvements/qa-rejection-log.md, agent-improvements/cost-calibration-log.md, .remember/recent.md, and feedback_* memories.
Boundary markers: None identified in the instructions for log reading.
Capability inventory: Can perform file writes via the Edit tool and execute local scripts.
Sanitization: Implements a human-in-the-loop requirement where all proposed edits must be approved by the user ('Ane') before execution.
[COMMAND_EXECUTION]: The skill executes a local maintenance script via python ~/.claude/hooks/maintenance_due.py. This is an expected operational command within the system's maintenance framework.
[REMOTE_CODE_EXECUTION]: The skill uses the ctx_execute_file tool to analyze agent-improvements/cost-calibration-log.md. This represents a dynamic execution pattern where a tool is run against a data file that may contain trace data from past user interactions.

improve-system