plannotator-compound

This skill conducts a comprehensive research analysis of a user's Plannotator plan archive to extract denial patterns, a feedback taxonomy, evolution over time, and actionable prompt improvements, then produces a polished HTML dashboard report. It is structured as a strict multi-phase process where each phase completes fully before the next begins and research integrity is paramount: every in-scope file must be read, with no skipping or early summarizing. The workflow first performs source selection. In first-class Plannotator mode it locates the data directory using $PLANNOTATOR_DATA_DIR or the default ~/.plannotator and looks for *-denied.md files in the plans/ subdirectory. If that archive is absent or has no denied plans, it falls back to Claude Code mode by reading ~/.claude/projects/ and using a bundled parser to extract ExitPlanMode denial reasons from JSONL transcripts. If neither source is available it asks the user rather than guessing.

Phase 0 locates the plans and detects previous reports, which follow a versioned naming scheme (compound-planning-report.html, then -v2, -v3, and so on). If a prior report exists, the latest one's filesystem modification date becomes a cutoff date and the user is offered an incremental run (only files dated after the cutoff, building on prior findings) or a full re-analysis, with the new report using the next version number either way. Phase 1 inventories the dataset, always counting all files for overall stats: approved and denied counts, the date range and days spanned, and a revision rate of denied over total. It deliberately ignores *.annotations.md files because denied files already contain the full plan plus appended reviewer feedback after a separator, and it runs wc -l over approved plans to track whether plan size is growing.

Phase 2 is the map step, reading every in-scope *-denied.md file and only those files. For incremental runs only files whose embedded YYYY-MM-DD date is strictly after the cutoff are processed. Extraction is parallelized through sub-agents using model haiku, since the work is straightforward file reading and structured extraction rather than reasoning, with a batching strategy scaled to dataset size: tiny datasets read directly in the main agent, small datasets use 2-3 agents, medium datasets 4-6, and large datasets as many as needed to keep batches around 10-15 files split along natural time boundaries. Use this skill to turn an archive of denied plans into an evidence-based, versioned dashboard of where planning went wrong and how to improve prompts.