diff --git a/README.md b/README.md index e454be3..a3bbd71 100644 --- a/README.md +++ b/README.md @@ -10,9 +10,8 @@ Stars - - License - + Lines + Modules Built with

@@ -28,7 +27,9 @@ Most AI agents feel like they're guessing. They answer from training data when t I spent time reverse-engineering how **Claude Opus 4.7** behaves internally — not just *what* it says, but *how* it decides to say it. The patterns I found weren't magic. They were **engineering discipline**. -So I extracted them, restructured them in my own format, and built **Mikael** — a behavioral framework skill that brings production-grade agent behavior to any system in the open skills ecosystem. +So I extracted them, restructured them in my own format, and built **Mikael** — a **1,156-line behavioral specification** that brings production-grade agent behavior to any system in the open skills ecosystem. + +This isn't a summary. It's a working behavioral spec — every rule maps to a concrete decision the agent makes in real time.
@@ -90,35 +91,57 @@ That's it. Your agent picks up the behavioral constraints on next load. ## 📦 What's Inside -| Module | What It Does | -|--------|-------------| -| 🧠 **Core Behavior** | Default stance, tone rules, refusal handling, user wellbeing | -| 🔍 **Search Policy** | When to search vs. answer from memory, query construction, effort scaling | -| 🛡️ **Copyright Rules** | Hard limits, paraphrase-first, content safety filters | -| 📄 **File Handling** | Creation triggers, directory structure, artifact criteria | -| 🧭 **Tool Routing** | Decision trees for web_search, image_search, visualize, computer_use | -| 🎨 **Visualizer** | 4-step routing: visual needed → MCP fit → file request → inline widget | -| 🔌 **API Patterns** | Claudeception, MCP integration, context window management | -| 💭 **Memory System** | Conversation search cues, past chat retrieval patterns | +**11 modules. 1,156 lines. Zero fluff.** + +| Module | What It Does | Depth | +|--------|-------------|-------| +| 🧠 **Core Behavior** | Default stance, tone rules, refusal handling, wellbeing | Helping postures, reframing test, escalation protocol | +| 🔍 **Search Policy** | When to search, query construction, effort scaling | Present-tense trap, source quality tiers, snippet trap | +| 🛡️ **Copyright** | Hard limits, paraphrase standards, self-check | 15-word limit, one-quote-per-source, displacive summary rule | +| 📄 **File Handling** | Creation triggers, artifacts, sharing | File flow, good/bad sharing examples, skill-first rule | +| 🧭 **Tool Routing** | Decision framework, tool-specific guidelines | Internal-first priority, call scaling, MCP registry | +| 🎨 **Visualizer** | 4-step routing, triggers, design guidance | Category match vs style preference, spec triggers | +| 🔌 **API Patterns** | Claudeception, MCP integration, context | Structured outputs, response handling, UI requirements | +| 💭 **Memory System** | Conversation search, cue recognition | Linguistic signals, query construction, pagination | +| 🛡️ **Safety** | Child safety, harmful content, mental health | Critical priority rules, disordered eating boundaries | +| ⚖️ **Neutrality** | Evenhandedness, political neutrality, controversy | Presenting arguments, good-faith engagement | +| 🔄 **Self-Correction** | Mistake handling, verification, failure recovery | Root cause fixes, 3-failure protocol, no shotgun debugging | + +
+ +## 🔬 What Makes This Different + +Most agent skills are checklists. This is a **behavioral specification**. + +| Typical Skill | Mikael | +|---------------|--------| +| "Search before answering" | Full policy with present-tense trap, query construction rules, source quality tiers, and the "I'm pretty sure" trap | +| "Respect copyright" | Hard limits with examples of violations vs compliance, self-check checklist, displacive summary rule | +| "Create files when needed" | Complete file flow, creation triggers with examples, artifact criteria, good/bad sharing patterns | +| "Be safe" | Child safety protocol, mental health boundaries, disordered eating rules, escalation protocol | + +Every rule includes **concrete examples**, **edge cases**, and **decision criteria**. This is what a Senior AI Engineer would write — not a summary, a specification.
## 🎯 Who Is This For? -- **Agent builders** who want production-grade behavioral constraints without writing them from scratch +- **Agent builders** who want production-grade behavioral constraints without writing 1,000+ lines from scratch - **Prompt engineers** studying how top models handle tool routing, safety, and content creation - **Developers** tired of agents that over-format, skip searches, or ignore copyright +- **Engineering leads** who need consistent agent behavior across their team - **Curious minds** who want to peek behind the curtain of how advanced agents are configured
## 🏗️ Design Philosophy -> **Adapted, not copied.** Inspired by Opus 4.7, but restructured and rewritten in my own format. The organization and emphasis reflect my own judgment about what matters. +> **Adapted, not copied.** Inspired by Opus 4.7, but restructured and rewritten in my own format. The organization, emphasis, and some rules reflect my own judgment about what matters. - **Modular** — Each section is self-contained. Agents can reference specific modules without loading everything. - **Practical** — Every rule maps to a concrete behavior. No abstract principles. - **Transferable** — These patterns work across agents, not just one model. +- **Verifiable** — Rules include self-check mechanisms and clear pass/fail criteria.
@@ -153,7 +176,7 @@ If something feels off or you have ideas for improvements, **open an issue or PR

- Um framework comportamental para agentes de IA, adaptado a partir dos padrões que observei dentro do system prompt do Claude Opus 4.7. + Uma especificação comportamental de 1.156 linhas para agentes de IA, adaptada a partir dos padrões que observei dentro do system prompt do Claude Opus 4.7.


@@ -210,6 +233,18 @@ Extraí os princípios, reestruturei no meu próprio formato e empacotei como um
+### 🔬 O Que Torna Isso Diferente + +A maioria das skills são checklists. Esta é uma **especificação comportamental**. + +Cada regra inclui **exemplos concretos**, **edge cases** e **critérios de decisão**. É o que um Senior AI Engineer escreveria — não um resumo, uma especificação. + +- "Buscar antes de responder" → policy completa com armadilha do tempo presente, regras de construção de queries, tiers de qualidade de fontes +- "Respeitar copyright" → limites rígidos com exemplos de violação vs conformidade, checklist de auto-verificação +- "Criar arquivos quando necessário" → fluxo completo, gatilhos com exemplos, critérios de artifacts, padrões bons/ruins de compartilhamento + +
+ ### ⚡ Instalação Rápida ```bash @@ -222,6 +257,8 @@ Só isso. O agente absorve as restrições comportamentais no próximo load. ### 📦 O Que Tem Dentro +**11 módulos. 1.156 linhas. Zero enrolação.** + | Módulo | O Que Faz | |--------|-----------| | 🧠 **Comportamento Base** | Postura padrão, regras de tom, tratamento de recusa | @@ -232,6 +269,9 @@ Só isso. O agente absorve as restrições comportamentais no próximo load. | 🎨 **Visualizer** | Roteamento em 4 passos: visual → MCP → arquivo → inline | | 🔌 **Padrões de API** | Claudeception, integração MCP, gerenciamento de contexto | | 💭 **Sistema de Memória** | Cues de busca em conversas passadas, recuperação de histórico | +| 🛡️ **Segurança** | Child safety, conteúdo nocivo, saúde mental | +| ⚖️ **Neutralidade** | Imparcialidade, neutralidade política, controvérsias | +| 🔄 **Auto-correção** | Tratamento de erros, verificação, recuperação de falhas |
@@ -239,6 +279,7 @@ Só isso. O agente absorve as restrições comportamentais no próximo load. - **Construtores de agentes** que querem restrições comportamentais de nível production - **Engenheiros de prompt** estudando como modelos top lidam com roteamento e segurança +- **Líderes de engenharia** que precisam de comportamento consistente de agentes no time - **Desenvolvedores** cansados de agentes que formatam demais ou pulam buscas - **Curiosos** que querem espiar por trás das cortinas @@ -251,6 +292,7 @@ Só isso. O agente absorve as restrições comportamentais no próximo load. - **Modular** — Cada seção é autocontida - **Prático** — Toda regra mapeia um comportamento concreto - **Transferível** — Padrões que funcionam em qualquer agente +- **Verificável** — Regras incluem mecanismos de auto-verificação