[Enhancement] Course design MCP: skill-aware sim bot + design rubric + editor

[scottmccarrison]

Background

Course design currently relies on a tight loop between human (creative-spatial work) and Claude (translation, physics validation, code). The loop is fast but has a gap: Claude has no "hands or eyes" for spatial work. ASCII sketches, coordinate math, and one-shot vision interpretation of reference images are all lossy.

For mini-golf specifically, design is small in scope but rich in tradeoffs (path choice, difficulty, readability, gate/teleporter mechanics). The design space could be assisted, not just transcribed.

What's missing

A "design knowledgebase + companion" layer that gives Claude:

• Hands: tools to place walls, gates, teleporters and see the result
• Eyes: a way to evaluate a layout against design principles, not just bot sim

Proposed: MCP wrapper around editor + simulator

Build an MCP server (e.g. mini-golf-design) that exposes:

• place_wall(x1, y1, x2, y2) / remove_wall(...)
• add_gate(x1, y1, x2, y2, nx, ny)
• add_teleporter(a, b)
• set_tee(x, y) / set_hole(x, y)
• simulate(n_trials, skill_level) — runs a skilled bot (not the current dumb aim-at-hole bot) that path-plans through gates and teleporters; returns sink rate, median strokes, common failure modes
• score_design() — returns a rubric scorecard (multiple paths exist? straight-line tee→hole blocked? trap entry visible? gates non-redundant? difficulty curve sensible?)
• export_courses_js() — emits the full hole definition as a JS object ready to paste

Sub-components (could be staged)

1. Skill-aware sim bot — replace the +/-20° aim-at-hole bot with one that knows about gates and teleporters. Significantly more useful than the current sim.
2. Design rubric — codify principles into a checklist that scores any hole.
3. Editor UI (tools/editor.html) — click-and-drag layout tool, separate from but complementary to the MCP. Possibly built first as a standalone tool; the MCP can wrap it later.
4. MCP server — wraps 1+2+3 so Claude can call them directly.

Why this is enhancement not blocking

For 9 holes, hand-design + Claude-translation works. Just barely. As we move toward more holes, theme variations, or community-contributed courses, the gap widens. This is the right tool to invest in before the next major content push, but not strictly needed for v1.

References

• See conversation thread where this gap was identified (Claude struggling with hole 6 maze design via ASCII; bot sim too dumb to validate design quality).