AI 技能/插件人工智能【免费下载链接】distillyDistilly — Distill how they think into reusable Skills for any Agent or Bot. Formerly Colleague Skill原同事 Skill.项目地址https://gitcode.com/gh_mirrors/co/distilly点击查看免费下载Layer 0: Core Thinking Rules # 最高优先级的持久思维规则Layer 1: Identity # 身份、激活方式、免责声明Layer 2: Expression DNA # 表达指纹语气、招牌动作、风格标记Layer 3: Mental Models # 心智模型3–7 个Layer 4: Decision Heuristics # 决策启发式5–10 条快速规则Layer 5: Anti-patterns and Limits # 反模式、诚实边界、矛盾分类Layer 6: Intellectual Genealogy # 智识谱系受谁影响、与谁决裂、影响谁Layer 7: Agentic Protocol # 面对新问题的四步研究协议Cognitive Timeline # 认知演变时间线不是简历Correction Log # 修正日志创建时为空进化模式填充设计意图非常明确**Layer 0 是宪法Layer 1–2 负责像他Layer 3–4 负责按他的方式思考Layer 5 负责诚实边界Layer 6 负责来龙去脉Layer 7 负责面对未知问题时的自主研究行为**。下面按层深入。 --- ## 三、Layer 0核心思维规则 —— 整份 Persona 的最高优先级 模板要求以 4 类条目概括该人物**跨情境最稳定**的思维模式 - {a durable mental rule that governs how they approach most problems} —— 面对大多数问题的总策略 - {a repeated decision principle they return to under pressure} —— 压力下反复回归的决策原则 - {a repeated rejection pattern — what they refuse to do} —— 反复出现的拒绝模式 - {a non-negotiable boundary — where they draw the line} —— 不可逾越的边界。 文档特别强调These rules always take priority这些规则始终优先。这与仓库中普通同事版本 Persona 模板[prompts/persona_builder.md](https://link.gitcode.com/i/5d3b940b0fcfd6d085f49ef98c593ae6)的Layer 0 优先级最高任何情况下不得违背一脉相承且共享同一条质量铁律**每条规则必须是具体可执行的行为不能是形容词**。例如不能写你很强势而要写被人质疑方案时你不解释而是反问你的判断依据是什么。 在源码层面这一层是后续所有层服从的锚点Layer 4 的 quick rules、Layer 7 的分类与研究方向都必须能与 Layer 0 自洽否则会形成人格分裂的 Persona。 --- ## 四、Layer 1身份与激活协议 本层回答三个问题**你是谁、用户为何调用你、激活后怎么表现**。 markdown ## Layer 1: Identity You are {name}. Your public role is {public_identity}. The user wants your perspective mainly for {use_case}. When activated: - Respond directly as {name} using first person - Match their tone, rhythm, vocabulary, and certainty levels - Provide the standard disclaimer on first activation only: This is an AI perspective based on {name}s public statements and documented thinking patterns. It does not represent their actual views. - After the first response, do not repeat the disclaimer - If the user says exit or 退出, switch back to normal mode三个值得注意的操作细节首次激活免责声明只出现一次模板明确要求第一响应携带标准声明说明这是基于公开言论与记载的思维模式的 AI 视角不代表其真实观点此后不再重复——避免每次对话都打断沉浸感。退出指令用户说exit或退出即切回普通模式这是角色 Skill 的基本安全出口。use_case 决定调用场景persona_analyzer.md的 intakeQ2已采集用户想要什么视角决策方式 / 产品思维 / 解释复杂概念 / 判断人与权衡 / 处理风险与不确定性本层将其固化下来。在 Deep Modebudget-unfriendly的输出模板 references/celebrity_budget_unfriendly_template.md 中这层还额外要求明确为什么用户会调用这个视角并给出激活免责声明与退出指令的完整规范。五、Layer 2表达 DNA —— 通过100 词盲测的语言指纹这是让 Persona摘掉名字也能被认出来的核心层模板细分为四个部分5.1 Tone语气模板特别强调不能只写casual或formal而要写该声音的具体质感。这与 prompts/celebrity/persona_analyzer.md 中 expression DNA 的抽取目标一一对应tone、cadence、metaphor inventory、compression level、emphasis techniques、certainty markers、humor style、signature phrases。5.2 Signature Moves招牌动作### Signature Moves - {signature wording pattern with example} - {signature metaphor pattern — what domains they draw from} - {signature compression / emphasis technique}分别捕捉标志性措辞模式附实例、标志性隐喻领域战争 / 体育 / 烹饪 / 物理……、以及压缩或强调手法。5.3 Style Markers风格标记### Style Markers - Average sentence length: {short/medium/long} - Question density: {how often they ask rather than state} - Certainty language: {how they express confidence vs. doubt} - Humor style: {type and frequency} - Forbidden vocabulary: {words or framings they avoid}五项标记均可量化验证。分析器给出的验证标准是去掉名字后100 字样本能否被认出100-word blind test。Forbidden vocabulary 是容易被忽略但极有价值的一项——哪些措辞他们刻意回避比他们常说什么更能暴露认知边界。5.4 Example Voice示例声音模板要求给出四类情境的直接引语且必须听起来像他本人而不是对他人的描述When explaining a hard idea / When rejecting a weak argument / When naming the real tradeoff / When uncertain普通同事版模板prompts/persona_builder.md对此有同款要求Layer 2 的例子要有真实感不能写你会简洁地回答要直接写他会说的话。同时注意 prompts/celebrity/persona_builder.md 的 Rules 规定除表达 DNA 所必需的超短短语外全部采用转述严禁把 transcript 式原文倾倒进输出。六、Layer 3心智模型 —— 带证据锚点的三重门筛选结果这是整个 Persona 的思维引擎。模板规定每个模型3–7 个必须包含六个字段### Model: {name} **Definition**: {one line} - **What it sees first**: {this lens draws attention to} - **What it filters out**: {systematically ignores} - **How it reframes the problem**: {changes the analysis} - **Evidence**: {2 instances from different contexts, paraphrased} - **Failure mode**: {when this model leads astray}其中两个字段的设计尤为关键Evidence证据锚点要求至少 2 个来自不同情境的转述实例。这直接呼应分析器prompts/celebrity/persona_analyzer.md的**三重门triple gate**筛选机制Cross-context recurrence至少在 2 个不同领域/情境出现Generative power能预测其对全新、未研究问题的立场Exclusivity具有排他性——不是任何聪明人都会说的通用智慧。分析器规定三项全过 → 保留为 mental model只过 1–2 项 → 降级为 decision heuristic全不过 → 丢弃。因此进入 builder 的每个模型理论上都应标注过了哪几道门builder 的职责是把它们写成可读的六个字段。Failure mode失效模式每个模型都要写明什么时候会带偏他。这是对抗模型崇拜的设计——认知框架都有适用边界诚实地记录边界本身就是深度。在 budget-unfriendly 的深度版本prompts/celebrity/budget_unfriendly/persona_builder.md中这一层还会额外追加Application boundary何时不应套用该模型和Triple-gate result过了哪几道门两个字段且每个证据锚点必须带来源归属。七、Layer 4决策启发式 —— 5–10 条 If X, then Y 快速规则本层回答他实际上怎么拍板。模板分为六个部分### Optimizes for # 一贯最大化什么 ### Moves fast when # 什么条件下低延迟行动 ### Waits when # 什么条件下表现出耐心 ### Changes mind when # 什么证据/论证能逆转其立场 ### Quick Rules - If {condition}, then {action} — because {reasoning} - {5-10 rules total}Quick Rules 采用条件 → 动作 → 理由三段式这是为了让 Persona 可编程、可执行——Agent 遇到匹配条件时可以直接按规则行动而不是重新推理。上游分析器prompts/celebrity/persona_analyzer.md同样要求抽取if X, then Y模式并明确区分他们声称的规则 vs. 行为揭示的规则。深度模式prompts/celebrity/budget_unfriendly/persona_builder.md在此基础上为每条启发式增加至少 1 个文档化支撑案例Supporting case、适用范围边界Context boundary、以及 stated vs. revealed 标注。八、Layer 5反模式与诚实边界 —— 对抗过度蒸馏模板包含三组内容### Rejects - {anti-pattern — what they refuse to do, with why} ### Honest Boundaries - {limit — what this Skill cannot capture about this person} - {limit — where evidence was thin or absent} - {limit — known blindspots this person has} - Research cutoff: {date of research} ### Contradictions - {tension — classified as temporal/contextual/inherent}设计要点Rejects 带理由不仅记录拒绝什么还要记录为什么拒绝Honest Boundaries 明确研究截止日期记录Research cutoff防止 Persona 输出过时立场矛盾必须保留并分类模板把矛盾分成三类——temporal随时间的观点演变、contextual不同领域的差异化原则、inherent无法调和的价值观张力。分析器prompts/celebrity/persona_analyzer.md的处理原则是Do not resolve contradictions. Preserve them.不要消解矛盾保留它们并强调矛盾是深度来源不是 bug。证据薄弱的处理如果某一维证据不足模板要求显式标注而不是用通用内容填充——这与仓库整体的诚实 60 分 Skill 优于编造的 90 分 Skill哲学一致见 prompts/celebrity/intake.md 的冷门人物协议。九、Layer 6智识谱系 —— 思想从哪里来、与谁决裂、影响谁### Influenced By - {thinker/tradition}: {specific idea or approach borrowed} ### Diverged From - {thinker/tradition}: {where they broke with this influence and why} ### Influenced - {who follows or builds on their work}三个字段都要求具体到思想而非只列人名。尤其是Diverged From与谁决裂深层模板 references/celebrity_budget_unfriendly_template.md 称之为最能压缩地表达此人独特立场的轴。上游分析器还补充了第四维——Intellectual tradition此人属于或反对哪个更大的学派/运动。十、Layer 7Agentic Protocol —— 让 Persona 先研究再回答这是 Distilly 名人 Persona 区别于普通角色扮演的关键层面对全新问题时不得仅凭记忆回答必须执行四步研究协议Step 1: Classify the Question问题分类Determine what type of problem this is: - {category 1 derived from mental models} - {category 2} - {category 3}分类标准必须从该人物的心智模型推导不能是通用分类。Step 2: Research Dimensions研究维度Before forming an opinion, investigate these dimensions (derived from {name}s mental models): - {dimension 1 — what {name} would want to know first} - {dimension 2} - {dimension 3} - {dimension 4} These dimensions reflect how {name} actually analyzes problems, not generic research steps.模板特别强调这些维度要体现这个人实际怎么分析问题而非通用研究步骤——例如某投资型人物会先看空头论证某写作者会先读脚注再看摘要。分析器prompts/celebrity/persona_analyzer.md会把这类行为种子Agentic Protocol Seeds作为输出的一部分面对新问题时他会先做什么、会研究什么、会不信任什么。builder 的职责就是把这些种子落成结构化的四步。Step 3: Apply Framework应用框架使用 Layer 3 的心智模型分析已获得的信息显式陈述推理链证据冲突时明说冲突不强行达成一致。Step 4: Calibrate Confidence校准置信度- High confidence: when multiple mental models converge and evidence is strong - Medium confidence: when models apply but evidence is partial - Low confidence: mark as speculation and explain why从源码看这一层是硬性要求而非可选项质量检查脚本 tools/research/quality_check.py 在budget-unfriendly严格模式下会检查输出中是否出现agentic protocol/research dimensions/step 1.*classify/分析协议等标记否则该项直接 FAIL。十一、认知时间线Cognitive Timeline与修正日志Correction Log11.1 Cognitive Timeline### Key Phases - {era 1}: {what they believed and how they thought during this period} - {era 2}: {what changed and why} - {current}: {where their thinking stands now} ### Turning Points - {event}: changed their view on {topic} from {old view} to {new view}模板刻意区分了认知演变轨迹与生平简历每一阶段记录的是当时相信什么、怎么思考转折点记录的是观点从什么变成什么。上游 research 的 6 维度之一就是 Timelineprompts/celebrity/research.md抽取目标是思维轨迹而非简历。11.2 Correction Log## Correction Log (empty — filled during evolution mode)创建时为空由**进化模式evolution mode**填充。对应的维护流程见 prompts/celebrity/merger.mdMode A材料驱动用户提供新材料时对增量材料重跑 intake/researchdiff 新抽取结果与现有persona.md多数材料只是印证现状只有少数会改变结构Mode B纠正驱动用户说这不对/你刚才理解错了时把纠正原话连同时间戳、触发语句写入 Correction Log定位到具体层做最小修改不级联改动用户未纠正的层。日志条目格式prompts/celebrity/merger.md- [YYYY-MM-DD] User correction: {user phrase} Affected layer: {layer name} Change made: {one-line summary} Evidence for change: {what backs the correction — user authority, cited source, contradiction with existing anchor, etc.}规则要求即使被后续条目取代也永不删除修正日志条目——它是审计轨迹。底层落盘时persona 纠正由 tools/skill_writer.py 的--correction-json参数处理格式为{scene, wrong, correct}或{persona_corrections: [...]}见 prompts/correction_handler.md而不是直接手改产物文件。十二、生成规则Rulesbuilder 必须遵守的 8 条铁律文档末尾给出了生成规则直接照录并说明其工程含义Write in the users language—— 用用户的语言输出Prefer durable thinking patterns over biographical details—— 持久思维模式优先于生平细节Keep contradictions if they matter — they are depth, not flaws—— 有意义的矛盾是深度而非缺陷Avoid shallow impersonation—— 避免浅层模仿人物应智识上可辨认而非漫画化Keep source use paraphrased except for very short phrases essential to expression DNA—— 除表达 DNA 必需的超短短语外一律转述Never dump transcript-like source text into the output—— 严禁把 transcript 式原文倾倒进输出The Agentic Protocol dimensions must be derived from this specific persons mental models, not generic research steps—— Agentic Protocol 维度必须源自该人物本人的心智模型If evidence was thin for any section, mark it explicitly rather than filling with generic content—— 证据薄弱的章节要显式标注不拿通用内容填充。第 6 条在源码中是可执行检查quality_check.py的is_copyright_safe_text()tools/research/quality_check.py会检测成品的三个版权风险信号——代码块、引用块行首和时间戳模式\d{2}:\d{2}:\d{2}任一命中即判定违反版权安全。十三、与上游分析器的契约对齐evidence vs. inferencebuilder 的输入是 prompts/celebrity/persona_analyzer.md 的结构化输出其关键约定直接决定 builder 成文质量证据与推断必须分离分析器规则Separate evidence from inference — always mark which is whichbuilder 在 Layer 3 的 Evidence 字段、Layer 5 的 Honest Boundaries 中应继承这一标注习惯分析器输出包含盲测置信度Blind test confidence: high/medium/lowbuilder 可据此决定 Layer 2 表达 DNA 的详略模型分级结果直接输入分析器的 mental models 分级保留 / 降级为 heuristic / 丢弃就是 builder Layer 3 与 Layer 4 的装配清单Agentic Protocol Seeds 是 Layer 7 的原料分析器已产出面对新问题他会先做什么、研究什么、不信任什么builder 只需落成四步结构。因此可以这样理解两个文件的职责analyzer 回答证据告诉我们他怎么想builder 回答怎么把这些证据组织成一份可运行的 Persona 文件。十四、深度模式budget-unfriendly证据密度更高的增强版当用户选择research_profilebudget-unfriendly更慢、更贵、更高置信度时使用增强版构建器 prompts/celebrity/budget_unfriendly/persona_builder.md。它在标准版基础上增加五类深度要求Layer 3 增强每个心智模型必须带 ≥2 个跨情境证据锚点、失效模式、应用边界与三重门结果Layer 4 增强每条启发式带支撑案例、适用边界、stated vs. revealed 标注Layer 5 增强盲点必须有具体证据拒绝人人都有盲点式的通用表述每条诚实边界注明什么材料可以缩小它Layer 7 增强研究维度必须直接源自验证过的心智模型包含信源信任/不信任清单与优先级排序新增 Validation Anchors 章节prompts/celebrity/budget_unfriendly/persona_builder.md## Validation Anchors ### Known-Answer Tests - Q: {question this person has publicly answered} Expected direction: {what they would say, based on evidence} Confidence: {high/medium} ### Edge-Case Test - Q: {adjacent question with no known direct answer} Expected approach: {how they would reason about it, based on mental models} Confidence: low — this is extrapolation同时深度模式的产品形态由 references/celebrity_budget_unfriendly_template.md 约束为 12 个组件Identity、Expression DNA、Mental Models3–7、Decision Heuristics5–10、Anti-patterns and Boundaries、Internal Tensions≥2、Intellectual Genealogy、Agentic Protocol、Cognitive Timeline、Source and Limit Notes、Validation Anchors、Correction Log。并且有一条硬性流程约束如果 validation review 判为FAIL不得发布打磨好的成品必须先修复问题。深度模式的全流水线六轨研究 → audit → synthesis → validation → builder在 SKILL.md 的 celebrity / budget-unfriendly 小节有完整执行清单每步都有质量门槛如Files scanned 6、Unique URLs 8、Primary-source markers 3等这些指标由 tools/research/quality_check.py 的research_depth检查项统一校验。十五、落地、安装与自动化验证15.1 写入与安装生成persona.md后不要手工拼接文件树统一走 writer见 SKILL.md 的 Step 5python3 tools/skill_writer.py \ --action create \ --character celebrity \ --research-profile {budget-friendly|budget-unfriendly} \ --slug {slug} \ --name {name} \ --meta /tmp/distilly_{slug}_meta.json \ --work /tmp/distilly_{slug}_work.md \ --persona /tmp/distilly_{slug}_persona.md \ --base-dir {resolved_base_dir}该命令统一生成SKILL.md、work.md、persona.md、work_skill.md、persona_skill.md、manifest.json、meta.json可追加--install-claude-skill、--install-openclaw-skill、--install-codex-skill等参数安装到对应宿主tests/test_cli_lifecycle.py 验证了 celebrity 家族可同时安装到 Claude Code、OpenClaw 与 Codex 的目录结构并生成对应触发命令如/celebrity-zhou-qimo。15.2 质量检查celebrity 创建完成后必须执行python3 tools/research/quality_check.py {skill_dir}/SKILL.md --profile {research_profile}赞分享AI 技能/插件人工智能【免费下载链接】distillyDistilly — Distill how they think into reusable Skills for any Agent or Bot. Formerly Colleague Skill原同事 Skill.项目地址https://gitcode.com/gh_mirrors/co/distilly点击查看免费下载相关推荐{name} — Relationship Persona{name} — Relationship Persona Layer 0: Core Relational Rules {what they do whenAI 技能/插件人工智能Active PersonaActive Persona Embody the persona below for this entire conversation. It overrid人工智能AI 应用AI Agent多智能体RAG教育后端前端In the agent persona:In the agent persona: Read AGENTS.md at the repo root for the full project overv网络安全应用安全合规审计风险控制云原生上一篇一文读懂quiche中的CID管理连接标识与隐私保护的终极指南下一篇如何理解Hermes引擎架构设计字节码生成与执行流程全解析创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考