Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Runtime debugging and live project inspection via godot-mcp. Use when headless tools cannot diagnose the problem: code compiles but runtime behavior is wrong, tests pass but the game does not work as expected, or you need to inspect live state (node tree, console output, rendering). Escalation trigger: a task has failed 2+ times through the headless path (headless-build, gdunit-driver), or the issue is inherently runtime-only (visual glitches, input not responding, physics behaving unexpectedly
.claude/skills/randallliuxin-mcp-driver/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-07 | ✗→✓ | ▲ Improved | -3% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -15% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -46% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 1% | 0% |
| case-11 | ✗→✓ | ▲ Improved | -30% | 0% |
$ARGUMENTS
MCP is the upgrade path when faster tools hit a wall. The fast loop handles most issues:
| Problem type | Fast path (no MCP) | |---|---| | Syntax / parse error | headless-build | | Unit test failure | gdunit-driver | | API lookup | godot-api |
Escalate to MCP when:
If none of these apply, use the fast path skill instead.
When you escalate to MCP, follow this loop. Max 3 iterations before asking the user for help.
1. RUN — run_project (optionally with a specific scene)
2. CAPTURE — get_debug_output → collect errors, warnings, prints
3. ANALYZE — parse output, identify root cause
4. FIX — modify code / scene / config to address the issue
5. VERIFY — stop_project → run_project again, get_debug_output
6. REPORT — summarize what was wrong, what was changed, and the outcomeEach step is detailed in references/workflow.md. Common runtime issue patterns are in references/debug-patterns.md.
MCP tools require a projectPath parameter — the directory containing project.godot. Determine it from the current working directory or ask the user if ambiguous:
bash# Quick check ls project.godot 2>/dev/null && echo "Project root: $(pwd)"
Full tool documentation: references/tools.md
| Tool | Purpose | |---|---| | run_project | Launch project in debug mode, capture output | | get_debug_output | Retrieve console output and errors from running project | | stop_project | Stop the running project, get final output |
| Tool | Purpose | |---|---| | get_project_info | Analyze project structure (scene/script/asset counts) | | get_godot_version | Check installed Godot version | | list_projects | Find Godot projects in a directory | | launch_editor | Open the Godot editor (for manual inspection) |
| Tool | Purpose | |---|---| | create_scene | Create new scene file | | add_node | Add node to existing scene | | load_sprite | Load texture into Sprite2D node | | save_scene | Save/resave scene file | | export_mesh_library | Export scene as MeshLibrary |
| Tool | Purpose | |---|---| | get_uid | Get UID for a specific file | | update_project_uids | Resave all resources to update UID references |
| Iteration | Action | |---|---| | 1 | Run → capture → analyze → apply most likely fix → verify | | 2 | Re-analyze output, try alternative hypothesis → verify | | 3 | Broaden investigation (check project structure, cross-reference with reviewer skills) → verify | | After 3 | Stop and ask the user. Report what was tried and what was observed. |
After MCP reveals the root cause, hand off to the appropriate skill for the fix:
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 4,848 | 5,296 | +9% | 1 | 1 | 0% | 183 | 1,284 | +602% | 0 | 0 | — |
case-02 | fail→fail | 4,440 | 4,349 | -2% | 1 | 1 | 0% | 234 | 1,270 | +443% | 0 | 0 | — |
case-03 | fail→fail | 5,229 | 4,147 | -21% | 1 | 1 | 0% | 247 | 1,245 | +404% | 0 | 0 | — |
case-04 | pass→fail | 6,706 | 4,094 | -39% | 1 | 1 | 0% | 1,228 | 1,451 | +18% | 0 | 0 | — |
case-05 | pass→pass | 9,379 | 5,445 | -42% | 1 | 1 | 0% | 1,514 | 1,893 | +25% | 0 | 0 | — |
case-06 | pass→pass | 4,145 | 4,788 | +16% | 1 | 1 | 0% | 661 | 1,784 | +170% | 0 | 0 | — |
case-07 | fail→pass | 10,694 | 3,335 | -69% | 1 | 1 | 0% | 1,479 | 1,433 | -3% | 0 | 0 | — |
case-08 | fail→pass | 13,629 | 4,978 | -63% | 1 | 1 | 0% | 2,177 | 1,861 | -15% | 0 | 0 | — |
case-09 | fail→pass | 13,628 | 2,024 | -85% | 1 | 1 | 0% | 2,437 | 1,309 | -46% | 0 | 0 | — |
case-10 | fail→pass | 7,807 | 2,116 | -73% | 1 | 1 | 0% | 1,281 | 1,294 | +1% | 0 | 0 | — |
case-11 | fail→pass | 10,629 | 2,481 | -77% | 1 | 1 | 0% | 1,974 | 1,372 | -30% | 0 | 0 | — |
case-12 | fail→pass | 4,649 | 4,266 | -8% | 1 | 1 | 0% | 969 | 1,708 | +76% | 0 | 0 | — |
case-13 | fail→pass | 8,155 | 2,997 | -63% | 1 | 1 | 0% | 1,722 | 1,538 | -11% | 0 | 0 | — |
case-14 | pass→pass | 6,998 | 2,284 | -67% | 1 | 1 | 0% | 1,240 | 1,402 | +13% | 0 | 0 | — |
case-15 | fail→pass | 4,719 | 2,116 | -55% | 1 | 1 | 0% | 656 | 1,318 | +101% | 0 | 0 | — |
case-16 | pass→pass | 7,411 | 1,765 | -76% | 1 | 1 | 0% | 1,099 | 1,226 | +12% | 0 | 0 | — |
case-17 | pass→pass | 5,517 | 1,918 | -65% | 1 | 1 | 0% | 738 | 1,228 | +66% | 0 | 0 | — |
case-18 | pass→fail | 11,577 | 1,915 | -83% | 1 | 1 | 0% | 1,539 | 1,252 | -19% | 0 | 0 | — |
case-19 | fail→pass | 13,307 | 4,666 | -65% | 1 | 1 | 0% | 1,833 | 1,624 | -11% | 0 | 0 | — |
case-20 | fail→pass | 12,127 | 3,723 | -69% | 1 | 1 | 0% | 1,813 | 1,591 | -12% | 0 | 0 | — |
case-21 | fail→pass | 3,945 | 1,757 | -55% | 1 | 1 | 0% | 577 | 1,185 | +105% | 0 | 0 | — |
case-22 | fail→pass | 8,921 | 3,596 | -60% | 1 | 1 | 0% | 1,508 | 1,525 | +1% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 19 counted toward the lift figure. The other 3 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +45 percentage points is the difference between those two pass rates over the 19 comparable cases. 2 cases got worse with the skill loaded, and they are included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.