The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands such as ls, stat, and wc to gather metadata about plan files, calculate statistics, and detect previous report versions for incremental analysis.
[COMMAND_EXECUTION]: It invokes a bundled Python script (extract_exit_plan_mode_outcomes.py) to normalize and extract human-authored feedback from Claude Code session transcripts.
[DATA_EXFILTRATION]: The skill accesses sensitive data including user interaction transcripts and planning history located in ~/.claude/projects/ and ~/.plannotator/plans/. This data is processed locally to generate analytical reports and is not transmitted outside the user's environment.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection (Category 8) as it processes untrusted plan content and logs.
Ingestion points: Processes markdown files from Plannotator archives and JSONL transcripts from Claude Code logs.
Boundary markers: Employs structural delimiters in extraction prompts but lacks specialized guardrails to prevent obedience to instructions embedded within user plans.
Capability inventory: Executes system commands, runs local Python scripts, spawns sub-agents, and modifies the local hooks directory.
Sanitization: No explicit sanitization or validation of the processed user feedback is performed before its inclusion in the generated reports or the improvement hook file.

plannotator-compound