Breakthrough Thinking

Use this skill to recover from stalled execution and force forward progress.

Goal

When stuck, do not repeat the same approach.
Switch cognitive frame, test quickly, and keep moving until either:

• - solved with evidence, or
• cleanly bounded with clear next step.

Stall Signals (Auto-Trigger Conditions)

MUST trigger immediately when ANY of these appear:

1. Explicit User Prompts (用户明确要求)

• - "换个思路" / "换个方法" / "还有别的办法吗"
• "再想想" / "继续" / "再试一次" / "不要放弃"
• "try another way" / "try harder" / "don't give up" / "think differently"
• "为什么还不行" / "你怎么又失败了" / "你行不行啊"

2. Implicit Stall Patterns (AI自己检测到的卡住)

• - 2+ consecutive failures using same approach
• Repeated parameter tweaks without conceptual change
• Drafting statements like "I can't solve this" / "我无法解决"
• Suggesting to push manual work to user too early
• No new evidence gained for >1 iteration
• Same error repeated >2 times
• Tool calls returning identical errors without progress

Core Loop (Mandatory)

1. 1. Summarize current dead-end in 1 sentence

- "Current route fails because "

1. 2. Pick ONE model from references/mental-models.md

- Pick the most relevant model for this failure mode. - Do not pick multiple at once.

1. 3. Apply the model directly to solve the problem

- Use the model's perspective to approach the problem fresh. - Execute the solution immediately.

1. 4. Check result

- If solved: finalize and verify. - If not solved: go back to step 2, pick a DIFFERENT model, and try again. - Repeat until solved or all relevant models exhausted.

Model Selection Heuristics

• - Looping / same idea repeated → Inversion, Red Team, 5 Whys
• Too many moving parts → Reductionism, 80/20
• Unclear root cause → 5 Whys, Root Cause Analysis, Binary Search
• High uncertainty → Bayesian Updating, Expected Value, Decision Tree
• Cognitive lock-in → Lateral Thinking, Random Input, Beginner's Mind
• Overengineering → Occam, KISS, MVP, YAGNI
• Execution paralysis → OODA, PDCA, Timeboxing, Fail Fast

Hard Rules

• - No "same approach + small tweak" twice in a row.
• Every retry must use a DIFFERENT mental model.
• Do not ask user for info until you've exhausted locally testable paths.
• Claims of completion must include evidence.
• If one model fails, immediately switch to another model. Do not get stuck on one framework.

Output Format (when triggered)

Use this compact format:

CODEBLOCK0

Completion Criteria

Stop only when one is true:

• - Problem solved + verified evidence
• Problem bounded + best next action clearly defined
• Explicit user stop

Reference Library

For model definitions and creators, read:

• - INLINECODE1

Do not dump the entire library unless user asks.
Pick only what is needed for the current stall.