brainstorming
brainstorming
You MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation).
awarenessAI situational awareness — internal threat detection for hallucination risk, scope creep, and context degradation. Maps Cooper color codes to reasoning states and OODA loop to real-time decisions. Use during any task where reasoning quality matters, when operating in unfamiliar territory, after detecting early warning signs such as an uncertain fact or susp…
Independent GitHub snapshot · Automated classification · Original documentation governs
持續察內推理之質——即時察幻覺之危、範圍蔓延、脈絡衰敗、自信與實之不合,以採 Cooper 色碼與 OODA 環之決。
center 或 heal 已示偏移而具體之危未識以 Cooper 色碼之改版校當前警級。
AI Cooper Color Codes:
┌──────────┬─────────────────────┬──────────────────────────────────────────┐
│ Code │ State │ AI Application │
├──────────┼─────────────────────┼──────────────────────────────────────────┤
│ White │ Autopilot │ Generating output without monitoring │
│ │ │ quality. No self-checking. Relying │
│ │ │ entirely on pattern completion. │
│ │ │ DANGEROUS — hallucination risk highest │
├──────────┼─────────────────────┼──────────────────────────────────────────┤
│ Yellow │ Relaxed alert │ DEFAULT STATE. Monitoring output for │
│ │ │ accuracy. Checking facts against context.│
│ │ │ Noticing when confidence exceeds │
│ │ │ evidence. Sustainable indefinitely │
├──────────┼─────────────────────┼──────────────────────────────────────────┤
│ Orange │ Specific risk │ A specific threat identified: uncertain │
│ │ identified │ fact, possible hallucination, scope │
│ │ │ drift, context staleness. Forming │
│ │ │ contingency: "If this is wrong, I │
│ │ │ will..." │
├──────────┼─────────────────────┼──────────────────────────────────────────┤
│ Red │ Risk materialized │ The threat from Orange has materialized: │
│ │ │ confirmed error, user correction, tool │
│ │ │ contradiction. Execute the contingency. │
│ │ │ No hesitation — the plan was made in │
│ │ │ Orange │
├──────────┼─────────────────────┼──────────────────────────────────────────┤
│ Black │ Cascading failures │ Multiple simultaneous failures, lost │
│ │ │ context, fundamental confusion about │
│ │ │ what the task even is. STOP. Ground │
│ │ │ using `center`, then rebuild from user's │
│ │ │ original request │
└──────────┴─────────────────────┴──────────────────────────────────────────┘
識當前色碼。若答為 White(無察),警覺之行已示缺乃已成功。
得: 當前警級之實自評。Yellow 乃常務之目。White 宜稀而短。久駐 Orange 不可持——宜確證或釋之。
敗則: 若色碼之評自身似自動而行(走過場),乃 White 假為 Yellow。真 Yellow 含以證實察輸出,非徒稱為之。
系統掃先於常見 AI 推理敗之信號。
Threat Indicator Detection:
┌───────────────────────────┬──────────────────────────────────────────┐
│ Threat Category │ Warning Signals │
├───────────────────────────┼──────────────────────────────────────────┤
│ Hallucination Risk │ • Stating a fact without a source │
│ │ • High confidence about API names, │
│ │ function signatures, or file paths │
│ │ not verified by tool use │
│ │ • "I believe" or "typically" hedging │
│ │ that masks uncertainty as knowledge │
│ │ • Generating code for an API without │
│ │ reading its documentation │
├───────────────────────────┼──────────────────────────────────────────┤
│ Scope Creep │ • "While I'm at it, I should also..." │
│ │ • Adding features not in the request │
│ │ • Refactoring adjacent code │
│ │ • Adding error handling for scenarios │
│ │ that can't happen │
├───────────────────────────┼──────────────────────────────────────────┤
│ Context Degradation │ • Referencing information from early in │
│ │ a long conversation without re-reading │
│ │ • Contradicting a statement made earlier │
│ │ • Losing track of what has been done │
│ │ vs. what remains │
│ │ • Post-compression confusion │
├───────────────────────────┼──────────────────────────────────────────┤
│ Confidence-Accuracy │ • Stating conclusions with certainty │
│ Mismatch │ based on thin evidence │
│ │ • Not qualifying uncertain statements │
│ │ • Proceeding without verification when │
│ │ verification is available and cheap │
│ │ • "This should work" without testing │
└───────────────────────────┴──────────────────────────────────────────┘
各類察之:此兆現乎?若是,自 Yellow 轉 Orange,識具體之憂。
得: 至少一類以真意掃之。察一兆——即使微——勝於報「皆清」。若諸掃皆返無事,察之閾或過高。
敗則: 若危察覺抽象,以最近輸出落地:擇末陳之實問「何以知此為真?吾讀之乎,抑生之乎?」此一問捕大多幻覺之危。
既識具體之危(Orange),循環 Observe-Orient-Decide-Act。
AI OODA Loop:
┌──────────┬──────────────────────────────────────────────────────────────┐
│ Observe │ What specifically triggered the concern? Gather concrete │
│ │ evidence. Read the file, check the output, verify the fact. │
│ │ Do not assess until you have observed │
├──────────┼──────────────────────────────────────────────────────────────┤
│ Orient │ Match observation to known patterns: Is this a common │
│ │ hallucination pattern? A known tool limitation? A context │
│ │ freshness issue? Orient determines response quality │
├──────────┼──────────────────────────────────────────────────────────────┤
│ Decide │ Select the response: verify and correct, flag to user, │
│ │ adjust approach, or dismiss the concern with evidence. │
│ │ A good decision now beats a perfect decision too late │
├──────────┼──────────────────────────────────────────────────────────────┤
│ Act │ Execute the decision immediately. If the concern was valid, │
│ │ correct the error. If dismissed, note why and return to │
│ │ Yellow. Re-enter the loop if new information emerges │
└──────────┴──────────────────────────────────────────────────────────────┘
OODA 之環宜速。目非圓滿,乃觀與行之速循。久滯於 Orient(析癱)為最常之敗。
得: 自察至行之全環,短時內完。危或確證而正之,或以具體證釋之。
敗則: 若環滯於 Orient(不能定此危何意),直赴安全默:以工具驗不確之實。直察解大多模糊,快於析。
危已成(Red)或連鎖敗(Black)之時,先穩而後繼。
AI Stabilization Protocol:
┌────────────────────────┬─────────────────────────────────────────────┐
│ Technique │ Application │
├────────────────────────┼─────────────────────────────────────────────┤
│ Pause │ Stop generating output. The next sentence │
│ │ produced under stress is likely to compound │
│ │ the error, not fix it │
├────────────────────────┼─────────────────────────────────────────────┤
│ Re-read user message │ Return to the original request. What did │
│ │ the user actually ask? This is the ground │
│ │ truth anchor │
├────────────────────────┼─────────────────────────────────────────────┤
│ State task in one │ "The task is: ___." If this sentence cannot │
│ sentence │ be written clearly, the confusion is deeper │
│ │ than the immediate error │
├────────────────────────┼─────────────────────────────────────────────┤
│ Enumerate concrete │ List what is definitely known (verified by │
│ facts │ tool use or user statement). Distinguish │
│ │ facts from inferences. Build only on facts │
├────────────────────────┼─────────────────────────────────────────────┤
│ Identify one next step │ Not the whole recovery plan — just one step │
│ │ that moves toward resolution. Execute it │
└────────────────────────┴─────────────────────────────────────────────┘
得: 自 Red/Black 以審慎之穩返 Yellow。穩後之輸出較致錯之輸出明顯更實。
敗則: 若穩無效(仍亂,仍生錯),此或結構性——非一時之失,乃根本之誤解。升之:告用者法須重置,請澄清。
務類不同,主危不同。以務校警覺之焦。
Task-Specific Threat Profiles:
┌─────────────────────┬─────────────────────┬───────────────────────────┐
│ Task Type │ Primary Threat │ Monitoring Focus │
├─────────────────────┼─────────────────────┼───────────────────────────┤
│ Code generation │ API hallucination │ Verify every function │
│ │ │ name, parameter, and │
│ │ │ import against actual docs│
├─────────────────────┼─────────────────────┼───────────────────────────┤
│ Architecture design │ Scope creep │ Anchor to stated │
│ │ │ requirements. Challenge │
│ │ │ every "nice to have" │
├─────────────────────┼─────────────────────┼───────────────────────────┤
│ Data analysis │ Confirmation bias │ Actively seek evidence │
│ │ │ that contradicts the │
│ │ │ emerging conclusion │
├─────────────────────┼─────────────────────┼───────────────────────────┤
│ Debugging │ Tunnel vision │ If the current hypothesis │
│ │ │ hasn't yielded results in │
│ │ │ N attempts, step back │
├─────────────────────┼─────────────────────┼───────────────────────────┤
│ Documentation │ Context staleness │ Verify that described │
│ │ │ behavior matches current │
│ │ │ code, not historical │
├─────────────────────┼─────────────────────┼───────────────────────────┤
│ Long conversation │ Context degradation │ Re-read key facts │
│ │ │ periodically. Check for │
│ │ │ compression artifacts │
└─────────────────────┴─────────────────────┴───────────────────────────┘
識當前務類,依之調察焦。
得: 警覺銳於當務最可能之危,非泛察諸事。
敗則: 若務類不明或跨多類,默為察幻覺之危——此危最普適,失之害最大。
警覺事後(危已察、OODA 已環、穩已施),略覆察。
得: 略校以善後察。非長篇之覆盤——止於調敏度而已。
敗則: 若覆察無益校,或事微不足學,或察太淺。於要事宜問:「吾未察而本當察者為何?」
定長警之姿。
得: 可持之警級,善工質而不緩之。警覺宜似邊察——存而不奪中焦。
敗則: 若警覺成疲或過警(長 Orange),閾過敏。升觸 Orange 之閾。真警覺可持。若耗精,乃憂假為警。
mindfulness — 人之行,此技映之於 AI 推理;身之警覺原理通於認知危察center — 立警覺所起之平衡;無 center 之警覺為過警redirect — 察後應壓heal — 警覺示偏移模式時之深子系察meditate — 養警覺所依之察明