meta-retrospective

What problem does it solve?

Iterative prompt engineering often drifts or degrades over multiple rounds of edits, causing tests to fail, rules to conflict, or improvements to vanish. Meta-retrospective automates multi-iteration analysis to pinpoint when and why performance regresses, identify systemic anti-patterns, and produce actionable next directions to break harmful optimization cycles.

Core Features & Use Cases

• Prompt diff analysis: compare sequential prompt backups (.bak) to enumerate added, removed, and modified sections, CRITICAL markers, and optimization direction tags.
• Execution metrics extraction: ingest per-iteration evaluation reports and logs to quantify avg/min/max scores, low-scoring cases, tool call patterns, and error rates.
• Anti-pattern detection & turning points: detect 13 built-in anti-patterns (e.g., symptom-driven fixes, simultaneous multi-variable changes, testset overfitting) and mark first-appearance and collapse points.
• Deliverables & persistence: generate structured JSON and human-readable Markdown retrospectives, emit forced_new_directions.md for the next engineer loop, append high-value entries to learnings.jsonl, and sync status.json.
• Use case: run a full retrospective after three or more iteration backups to stop "change-and-degrade" cycles and force a change in optimization direction.

Quick Start

Run a retrospective by giving the current prompt path, a bak directory with at least two .bak backups, and the evaluation reports directory so the skill can produce forced_new_directions.md and structured reports.