coach

name	coach
description	Coach yourself — honest feedback on how you work with coding agents: prompt quality, recurring failure patterns, and what your work has been pointing at. From your own OpenStory store. Use when the user asks "how am I doing", "feedback on my prompting", "where do I get stuck", "what am I focused on lately", or "coach me".

coach

Hold up the mirror: what does the record say about how the user works. Pull signals, then assess honestly — anchored in data, not flattery.

Phase 1 — pull signals

Prefer a dedicated scorecard tool if it exists:

• mcp__openstory__prompt_scorecard — if available, use it (prompt-length distribution, one-shot rate, turns/session, edit-thrash). This tool may not exist yet — if not, compose the signals below.

Fallback (existing tools, all mcp__openstory__*):

• mcp__openstory__list_sessions — volume + session labels (for themes).
• mcp__openstory__session_patterns / mcp__openstory__session_errors — failure signals (loops, errors, recovery) on recent sessions.
• mcp__openstory__productivity — when the user works.
• mcp__openstory__tool_journey on a few heavy sessions — repeated identical tool calls are an edit-thrash signal.

Phase 2 — render

How you work — <window>
Strengths:
  • <a strength tied to a measurable signal>
Watch-outs:
  • <a recurring failure pattern + a concrete fix>
Direction:
  • <themes your recent work clusters into, from session labels>

Anchor every claim to a signal you actually pulled, and name it ("error rate N%", "M sessions on "). If you could only estimate (no scorecard tool), label it an estimate. Be direct and useful, not flattering.

Note: the deepest metrics (precise prompt-length distribution, edit-thrash rate) want the server-side prompt_scorecard MCP tool. Until it ships, this skill gives a solid heuristic read from patterns, errors, and labels.