The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes human-authored feedback and denial reasons extracted from planning archives (*-denied.md) and session logs (~/.claude/projects/*.jsonl). This unstructured text is interpolated into prompts for extraction and reduction agents without explicit sanitization, which could allow instructions embedded in the logs to influence sub-agent behavior.
Ingestion points: Accesses files in ~/.plannotator/plans/ and conversation transcripts in ~/.claude/projects/.
Boundary markers: Relies on markdown delimiters (---) to separate content types but does not provide specific guardrails to the LLM for ignoring instructions within the data.
Capability inventory: Spawns sub-agents (agent tool), performs filesystem writes, and executes bundled scripts.
Sanitization: No sanitization or escaping of the ingested text was observed.
[DATA_EXFILTRATION]: The skill accesses sensitive session logs in ~/.claude/projects/ to extract reasoning for plan denials. These transcripts contain full conversation history, which may include private code, configuration, or user data.
Evidence: Phase 0 and the log parser script target these directories to assemble the source dataset.
Context: This access is the primary data source for the skill's analytical functions.
[PERSISTENCE]: The skill offers to install an 'improvement hook' by writing prompt instructions to ~/.plannotator/hooks/compound/enterplanmode-improve-hook.txt. This file is designed to be automatically injected into future agent sessions to modify planning behavior.
Evidence: Phase 6 describes the creation and management of this persistent configuration file.
[EXTERNAL_DOWNLOADS]: The generated HTML dashboard template references external resources from a well-known service for typography.
Evidence: assets/report-template.html loads several font families from fonts.googleapis.com.
[COMMAND_EXECUTION]: The skill executes a bundled Python script to parse session transcripts.
Evidence: SKILL.md and references/claude-code-fallback.md provide instructions to execute scripts/extract_exit_plan_mode_outcomes.py via python3 to extract denial reasons.

plannotator-compound