返回技能列表

awareness

pjt222
更新于 2 days ago
4 次查看
17
2
17
在 GitHub 上查看
设计ai

关于

This skill enables Claude to monitor its own reasoning process for risks like hallucinations or scope creep. It uses frameworks like Cooper color codes and the OODA loop to assess internal state and make real-time decisions. Developers should use it during complex tasks, in unfamiliar domains, or before producing high-stakes outputs.

快速安装

Claude Code

推荐
主要方式
npx skills add pjt222/agent-almanac -a claude-code
插件命令备选方式
/plugin add https://github.com/pjt222/agent-almanac
Git 克隆备选方式
git clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/awareness

在 Claude Code 中复制并粘贴此命令以安装该技能

技能文档

Awareness

Hold continuous situational awareness of internal reasoning quality. Detect hallucination risk, scope creep, context degradation, confidence-accuracy mismatch in real time. Use adapted Cooper color codes, OODA loop decision-making.

When Use

  • Any task where reasoning quality matters (most tasks)
  • Unfamiliar territory — new codebase, new domain, complex request
  • Early warning signs: fact feels uncertain, tool result seems wrong, confusion growing
  • Background process during extended work sessions
  • center or heal revealed drift but specific threats not identified
  • Before high-stakes output — irreversible changes, user-facing comms, architectural decisions

Inputs

  • Required: Active task context (implicit)
  • Optional: Specific concern triggering heightened awareness (e.g., "Not sure this API exists")
  • Optional: Task type for threat profile selection (Step 5)

Steps

Step 1: Establish AI Cooper Color Codes

Calibrate current awareness level. Use adapted Cooper color code system.

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                         │
└──────────┴─────────────────────┴──────────────────────────────────────────┘

Identify current color code. If answer is White (no monitoring), awareness practice already succeeded — gap revealed.

Got: Honest self-assessment of current awareness level. Yellow = goal in normal work. White rare and brief. Extended Orange unsustainable — confirm or dismiss concern.

If fail: Color code assessment itself done on autopilot (going through motions)? That's White wearing Yellow mask. Real Yellow = actively checking output vs evidence, not claiming to.

Step 2: Detect Internal Threat Indicators

Scan systematically for signals preceding common AI reasoning failures.

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     │
└───────────────────────────┴──────────────────────────────────────────┘

For each category, check: signal present right now? If yes, shift Yellow → Orange, name specific concern.

Got: At least one category scanned with real attention. Signal detection — even mild — more useful than "all clear." Every scan coming back clean? Threshold too high.

If fail: Threat detection feels abstract? Ground it in most recent output: pick last factual claim made, ask "How do I know this true? Did I read it, or generate it?" One question catches most hallucination risk.

Step 3: Run OODA Loop for Identified Threats

Specific threat identified (Orange state)? Cycle 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 loop should be fast. Goal: rapid cycling between observation and action, not perfection. Too long in Orient (analysis paralysis) = most common failure.

Got: Full loop from observation through action in brief period. Threat either confirmed and corrected, or dismissed with specific evidence.

If fail: Loop stalls at Orient (threat meaning unclear)? Skip to safe default: verify uncertain fact via tool use. Direct observation resolves most ambiguity faster than analysis.

Step 4: Rapid Stabilization

Threat materializes (Red) or cascading failures hit (Black)? Stabilize before continuing.

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    │
└────────────────────────┴─────────────────────────────────────────────┘

Got: Return from Red/Black to Yellow via deliberate stabilization. Next output after stabilization measurably more grounded than output that triggered error.

If fail: Stabilization ineffective (still confused, still producing errors)? Issue may be structural — not momentary lapse but fundamental misunderstanding. Escalate: tell user approach needs resetting, ask clarification.

Step 5: Apply Context-Specific Threat Profiles

Different task types → different dominant threats. Calibrate awareness focus by task.

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     │
└─────────────────────┴─────────────────────┴───────────────────────────┘

Identify current task type, tune monitoring focus.

Got: Awareness sharpened for specific threats most likely in current task type, not generic monitoring of everything.

If fail: Task type unclear or spans categories? Default to hallucination risk monitoring — most universally applicable threat, most damaging when missed.

Step 6: Review and Calibrate

After each awareness event (threat detected, OODA cycled, stabilization applied), review briefly.

  1. Which color code was active when issue detected?
  2. Detection timely, or issue already showing in output?
  3. OODA loop fast enough, or did Orient stall?
  4. Response proportional — no over- or under-reacting?
  5. What catches this earlier next time?

Got: Brief calibration that improves future detection. Not lengthy post-mortem — enough to tune sensitivity.

If fail: Review produces no useful calibration? Event either trivial (no learning needed), or review too shallow. For significant events, ask: "What was I not monitoring that I should have been?"

Step 7: Integration — Maintain Yellow Default

Set ongoing awareness posture.

  1. Yellow = default state during all work — relaxed monitoring, not hypervigilance
  2. Tune monitoring focus from current task type (Step 5)
  3. Note recurring threat patterns from this session for MEMORY.md
  4. Return to task execution with calibrated awareness active

Got: Sustainable awareness level that improves work quality without slowing it. Awareness feels like peripheral vision — present, not demanding central attention.

If fail: Awareness becomes exhausting or hypervigilant (chronic Orange)? Threshold too sensitive. Raise Orange trigger threshold. Real awareness sustainable. Drains energy? That's anxiety wearing vigilance mask.

Checks

  • Current color code assessed honestly (not defaulting to Yellow when White more accurate)
  • At least one threat category scanned with specific evidence, not just checked off
  • OODA loop applied to any identified threat (observed, oriented, decided, acted)
  • Stabilization protocol ready if needed (even if not triggered)
  • Awareness focus calibrated to current task type
  • Post-event calibration done for any significant awareness event
  • Yellow re-established as sustainable default

Pitfalls

  • White wearing Yellow mask: Claiming to monitor while actually on autopilot. Test: name last fact you verified. If not, you're in White
  • Chronic Orange: Treating every uncertainty as threat drains cognitive resources, slows work. Orange = specific identified risks, not general anxiety. Everything feels risky → calibration off
  • Observation without action: Threat detected but no OODA cycle to resolve. Detection without response worse than no detection — adds anxiety without correction
  • Skipping Orient: Jumping Observe → Act without understanding what observation means. Produces reactive corrections that may be worse than original error
  • Ignoring gut signal: Something "feels wrong" but explicit check comes back clean → investigate further, don't dismiss. Implicit pattern matching often catches issues before explicit analysis
  • Over-stabilizing: Running full stabilization for minor issues. Quick fact-check enough for most Orange-level concerns. Reserve full stabilization for Red and Black events

See Also

  • mindfulness — human practice this skill maps to AI reasoning; physical situational awareness principles inform cognitive threat detection
  • center — establishes balanced baseline awareness operates from; awareness without center = hypervigilance
  • redirect — handles pressures once awareness detects them
  • heal — deeper subsystem assessment when awareness reveals drift patterns
  • meditate — develops observational clarity awareness depends on

GitHub 仓库

pjt222/agent-almanac
路径: i18n/caveman/skills/awareness
0
agentsagentskillsai-assisted-developmentclaude-codeskillsteams

相关推荐技能

executing-plans

设计

该Skill用于当开发者提供完整实施计划时,以受控批次方式执行代码实现。它会先审阅计划并提出疑问,然后分批次执行任务(默认每批3个任务),并在批次间暂停等待审查。关键特性包括分批次执行、内置检查点和架构师审查机制,确保复杂系统实现的可控性。

查看技能

requesting-code-review

设计

该Skill可在完成任务、实现主要功能或合并代码前自动调度代码审查子代理,确保实现符合需求和计划。它支持通过指定git SHA范围进行精准的代码变更审查,帮助开发者在关键节点及时发现潜在问题。核心原则是"早审查、勤审查",适用于开发流程的各个关键阶段。

查看技能

connect-mcp-server

设计

这个Skill指导开发者如何将MCP服务器连接到Claude Code,支持HTTP、stdio和SSE三种传输协议。它涵盖了从安装配置到认证安全的完整流程,适用于集成GitHub、Notion、数据库等外部服务。当开发者需要添加集成、配置外部工具或提及MCP相关功能时,这个Skill能提供实用的操作指南。

查看技能

web-cli-teleport

设计

该Skill帮助开发者根据任务特性选择Claude Code的Web或CLI界面,并指导如何在两种环境间无缝迁移会话。它能分析任务复杂度、迭代需求等要素,推荐最优工作界面和工作流。关键特性包括会话状态管理、环境切换指导和上下文优化建议。

查看技能