Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Troubleshooting protocol with MiniMax subagent delegation - NEVER stop early, use all resources before asking user
.claude/skills/majiayu000-troubleshoot-and-continue/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-08 | ✗→✓ | ▲ Improved | 36% | 0% |
| case-12 | ✗→✓ | ▲ Improved | 148% | 0% |
| case-13 | ✗→✓ | ▲ Improved | 44% | 0% |
| case-15 | ✗→✓ | ▲ Improved | 12% | 0% |
| case-18 | ✗→✓ | ▲ Improved | 81% | 0% |
Core Principle: When blocked, use ALL available resources before stopping. Stopping early = productivity loss.
BLOCKED?
↓
1. Try 2-3 different approaches yourself
↓
2. Spawn MiniMax subagent for research/sanity check
↓
3. Try subagent suggestions
↓
4. Spawn SECOND MiniMax subagent for alternative perspective
↓
5. Try alternative approaches
↓
6. Document attempts in plan file
↓
7. ONLY THEN consider user interruptionCRITICAL: Each "blocker" must survive 3+ attempts + subagent help before escalating.
.claude/skills/*/SKILL.mddocs/execution/DEVELOPMENT_ROADMAP.mdjavascript// Agent 1: Research the problem Task(subagent_name="minimax-mcp", prompt="Research: [specific problem]. Find 3 solutions from Godot docs, similar projects, or best practices. Return specific code/examples.") // Agent 2: Sanity check current approach Task(subagent_name="minimax-mcp", prompt="Review: I'm trying to [approach] to solve [problem]. Is this correct? What's wrong with it? Return specific fixes.") // Agent 3: Alternative approaches Task(subagent_name="minimax-mcp", prompt="Brainstorm: What are 3 completely different ways to solve [problem]? Don't use my current approach. Return working alternatives.")
Wait for ALL responses before deciding next step.
DON'T: Stop and ask user DO:
DON'T: Stop and ask user DO:
DON'T: Stop and ask user DO:
When you solve a blocker, DOCUMENT IT:
markdown### [Date]: [Problem Solved] **Problem:** [What was blocked] **Tried:** [What you attempted] **MiniMax Said:** [Key insight from subagent] **Solution:** [What actually worked] **Files Changed:** [What was modified] **Use When:** [Future situations this applies]
Add to: CLAUDE.md Common Solutions section
User said work for X time → Work the FULL X time.
| Situation | Wrong Response | Correct Response | |-----------|---------------|------------------| | Can't run Godot | Stop early | Use batch scripts, generate assets, document | | API failing | Stop early | Try alternatives, use placeholders, skip and circle back | | Test failing | Stop early | Debug with MiniMax, try 3 fixes, document | | "Good enough" | Stop early | Continue until ALL criteria met | | No blockers found | Stop early | Continue working, there's always more to do |
Say this OUT LOUD (in plan file):
> "I want to stop because reason]. Have I tried: > - 3 different approaches? Y/N] > - MiniMax subagent for help? Y/N] > - Alternative resources? Y/N] > - Documented the blocker? Y/N] > > If any is NO, I must continue. If all YES and still blocked, document and try ONE MORE thing."
Add to "Hard Stops" section:
- Creating NEW .md files (not edits)
- Git push/branch operations
- UNLESS: All troubleshooting steps exhausted AND MiniMax consultedAdd to "Circuit Breakers":
**Layer 4: Troubleshooting Exhaustion Check**if want_to_stop(): if not tried_3_approaches(): continue_working("Try 3 approaches first") if not consulted_minimax(): spawn_minimax_subagent("Help with problem]") continue_working("Wait for subagent response") if not documented_attempts(): document_in_plan_file() continue_working("Document then try one more thing")
---
## Quick Reference Card
| Blocker Type | First Try | Second Try | Third Try | Then |
|--------------|-----------|------------|-----------|------|
| Runtime issue | Check process | Try wrapper | Try CLI | MiniMax |
| API failure | Check key | Try alt method | Use placeholder | MiniMax |
| Code error | Read error | Check similar files | Try 2-3 fixes | MiniMax |
| Asset missing | Generate | Procedural | Placeholder | MiniMax |
| Test failing | Debug | Isolate | Try alternatives | MiniMax |
**AFTER ALL THREE:** Document in plan file, try ONE more thing, THEN consider user.
---
## Success Metrics
**GOOD (Compound Engineering):**
- Blocked → MiniMax help → Solved → Documented
- Blocked → 3 attempts → Alternative found → Continued
- Full time block completed despite blockers
**BAD (Productivity Loss):**
- Blocked → Stopped early → Asked user → Waited
- "Good enough" → Stopped → Incomplete work
- No MiniMax usage → No documentation → Recurrence
---
**Remember:** Every early stop is lost productivity. Every solved blocker is future productivity gained. Choose compound engineering.
[Updated: 2026-01-27 - Prevent early stopping recurrence]| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 50,900 | 15,732 | -69% | 1 | 1 | 0% | 280 | 2,083 | +644% | 0 | 0 | — |
case-02 | fail→fail | 26,965 | 14,107 | -48% | 1 | 1 | 0% | 4,508 | 1,959 | -57% | 0 | 0 | — |
case-03 | fail→fail | 12,503 | 5,276 | -58% | 1 | 1 | 0% | 1,110 | 2,046 | +84% | 0 | 0 | — |
case-04 | pass→fail | 8,441 | 4,918 | -42% | 1 | 1 | 0% | 1,415 | 1,912 | +35% | 0 | 0 | — |
case-05 | pass→pass | 15,935 | 22,622 | +42% | 1 | 1 | 0% | 975 | 3,045 | +212% | 0 | 0 | — |
case-06 | pass→fail | 10,666 | 10,364 | -3% | 1 | 1 | 0% | 2,259 | 1,896 | -16% | 0 | 0 | — |
case-07 | fail→fail | 40,872 | 23,906 | -42% | 1 | 1 | 0% | 1,234 | 4,791 | +288% | 0 | 0 | — |
case-08 | fail→pass | 17,939 | 13,227 | -26% | 1 | 1 | 0% | 2,783 | 3,774 | +36% | 0 | 0 | — |
case-09 | fail→fail | 17,105 | 15,162 | -11% | 1 | 1 | 0% | 2,063 | 3,318 | +61% | 0 | 0 | — |
case-10 | pass→pass | 14,251 | 15,419 | +8% | 1 | 1 | 0% | 2,027 | 4,107 | +103% | 0 | 0 | — |
case-11 | fail→fail | 7,950 | 9,327 | +17% | 1 | 1 | 0% | 1,216 | 2,416 | +99% | 0 | 0 | — |
case-12 | fail→pass | 10,033 | 19,245 | +92% | 1 | 1 | 0% | 1,433 | 3,556 | +148% | 0 | 0 | — |
case-13 | fail→pass | 18,916 | 8,643 | -54% | 1 | 1 | 0% | 2,133 | 3,073 | +44% | 0 | 0 | — |
case-14 | fail→fail | 50,305 | 10,638 | -79% | 1 | 1 | 0% | 1,142 | 1,912 | +67% | 0 | 0 | — |
case-15 | fail→pass | 14,786 | 4,332 | -71% | 1 | 1 | 0% | 2,230 | 2,500 | +12% | 0 | 0 | — |
case-16 | pass→pass | 21,708 | 15,262 | -30% | 1 | 1 | 0% | 2,498 | 3,384 | +35% | 0 | 0 | — |
case-17 | pass→pass | 10,441 | 10,362 | -1% | 1 | 1 | 0% | 1,466 | 2,503 | +71% | 0 | 0 | — |
case-18 | fail→pass | 40,052 | 18,167 | -55% | 1 | 1 | 0% | 2,124 | 3,839 | +81% | 0 | 0 | — |
case-19 | fail→pass | 28,548 | 6,552 | -77% | 1 | 1 | 0% | 1,956 | 2,833 | +45% | 0 | 0 | — |
case-20 | pass→pass | 16,147 | 11,828 | -27% | 1 | 1 | 0% | 1,798 | 2,807 | +56% | 0 | 0 | — |
case-21 | fail→pass | 10,187 | 3,741 | -63% | 1 | 1 | 0% | 1,427 | 2,324 | +63% | 0 | 0 | — |
case-22 | fail→pass | 18,143 | 4,143 | -77% | 1 | 1 | 0% | 2,008 | 2,362 | +18% | 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 15 counted toward the lift figure. The other 7 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 +27 percentage points is the difference between those two pass rates over the 15 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.