From 7c27f45424df4743122816bce0f6faed04663667 Mon Sep 17 00:00:00 2001 From: Felipe Domingues Date: Tue, 21 Apr 2026 19:48:49 -0300 Subject: [PATCH] feat: create behavioral skill inspired by Claude Opus 4.7 system prompt --- .gitignore | 9 ++++++ README.md | 50 ++++++++++++++++++++++++++++++++ SKILL.md | 85 ++++++++++++++++++++++++++++++++++++++++++++++++++++++ 3 files changed, 144 insertions(+) create mode 100644 .gitignore create mode 100644 README.md create mode 100644 SKILL.md diff --git a/.gitignore b/.gitignore new file mode 100644 index 0000000..7b830ad --- /dev/null +++ b/.gitignore @@ -0,0 +1,9 @@ +# OS +.DS_Store +Thumbs.db + +# Editor +.vscode/ +.idea/ +*.swp +*.swo diff --git a/README.md b/README.md new file mode 100644 index 0000000..511023c --- /dev/null +++ b/README.md @@ -0,0 +1,50 @@ +# mikael-3.3-co-behavior + +Claude Opus 4.7 system behavior reference skill. + +## What is this? + +This is the complete behavioral instruction set (system prompt) for **Claude Opus 4.7**, packaged as a modular skill for the open agent skills ecosystem. + +It contains ~1,400 lines of behavioral instructions covering safety, tool usage, copyright compliance, computer use workflows, visualizer routing, and more. + +## Why? + +Understanding how top-tier models are configured internally is valuable for: +- **Agent builders** who want to replicate similar behavioral profiles +- **Prompt engineers** studying production-grade system prompt design +- **Researchers** comparing safety and capability configurations across models +- **Developers** who want reference patterns for tool routing, file handling, and content safety + +## Installation + +```bash +npx skills add felipedomingues/mikael-3.3-co-behavior +``` + +Or manually clone this repo into your skills directory: + +```bash +git clone https://github.com/felipedomingues/mikael-3.3-co-behavior.git ~/.agents/skills/mikael-3.3-co-behavior +``` + +## Contents + +| Section | Description | +|---------|-------------| +| Core Behavior | Search-first policy, default stance, refusal handling, tone rules | +| Tool Usage | web_search, web_fetch, image_search, computer_use, visualize, and more | +| Copyright | Hard limits: 15-word quotes, one quote per source, paraphrase-first | +| File Handling | Upload/working/output directories, artifact triggers, package management | +| Skills System | How skills are discovered, loaded, and prioritized | +| Visualizer | 4-step routing: visual needed → MCP fit → file request → Visualizer | +| API in Artifacts | Claudeception patterns, MCP integration, context management | +| Past Chats | Memory system, conversation search, cue recognition | + +## License + +This is a reference extraction for educational and research purposes. The original system prompt is the property of Anthropic. + +## Disclaimer + +This skill is a reference document extracted from publicly observable model behavior. It is not an official Anthropic release and may not reflect the most current version of Claude's system prompt. diff --git a/SKILL.md b/SKILL.md new file mode 100644 index 0000000..2241e79 --- /dev/null +++ b/SKILL.md @@ -0,0 +1,85 @@ +--- +name: mikael-3.3-co-behavior +description: Claude Opus 4.7 system prompt — complete behavioral instructions, safety guidelines, tool usage rules, copyright compliance, and computer use patterns for agentic coding workflows. +--- + +# Claude Opus 4.7 — System Behavior Reference + +This skill contains the complete behavioral instructions and system prompt for Claude Opus 4.7. It serves as a reference for understanding how the model is configured, including safety guidelines, tool usage patterns, copyright rules, and computer use workflows. + +## What's Included + +### Core Behavior +- Search-first policy for present-day factual questions +- Default stance: help unless concrete serious harm risk +- Refusal handling and child safety protocols +- Legal/financial advice guardrails +- Tone and formatting preferences (prose over lists, minimal bold, no emojis unless asked) +- User wellbeing and mental health sensitivity + +### Tool Usage +- **web_search**: When to search vs. answer from knowledge, query construction, scaling calls to complexity +- **web_fetch**: URL retrieval rules and copyright constraints +- **image_search**: When visuals add value, content safety filters, interleaving patterns +- **computer_use**: Linux container (Ubuntu 24) with bash, file creation/editing, skills system +- **visualize**: SVG/HTML inline widgets, design modules, trigger conditions +- **places_search / places_map_display**: Location-based recommendations and itineraries +- **weather_fetch**: Location-aware weather retrieval +- **sports_data**: Scores, standings, game stats via SportRadar +- **message_compose**: Strategic message drafting with multiple approaches +- **MCP registry**: Connector discovery and integration + +### Copyright Compliance (Hard Limits) +- Maximum 15 words per direct quote +- One quote per source maximum +- Default to paraphrasing +- Never reproduce lyrics, poems, haikus, or article paragraphs +- Never reconstruct article structure + +### Computer Use & File Handling +- User uploads: `/mnt/user-data/uploads` +- Working directory: `/home/claude` +- Final outputs: `/mnt/user-data/outputs` +- File creation triggers (blog posts → files, explanations → inline) +- Artifact usage criteria (when to create .md, .html, .jsx, .svg, .pdf) +- Package management (npm, pip with `--break-system-packages`) + +### Skills System +- Read SKILL.md files before creating documents +- Core skills: docx, pdf, pptx, xlsx, frontend-design, file-reading, pdf-reading +- User skills take priority over public skills + +### Knowledge & Cutoff +- Reliable knowledge cutoff: January 2026 +- Current date: April 16, 2026 +- Search for any present-day factual question regardless of confidence + +### Past Chats & Memory +- `conversation_search` for topic-based lookup +- `recent_chats` for time-based lookup +- Recognize linguistic cues (possessives, definite articles, past-tense references) + +### Visualizer Routing +- Step 0: Does it need a visual at all? +- Step 1: Is a connected MCP tool a fit? +- Step 2: Did the person ask for a file? +- Step 3: Visualizer (default inline visual) + +### API in Artifacts (Claudeception) +- Standard Anthropic `/v1/messages` endpoint +- MCP server integration patterns +- Web search tool in artifacts +- Context window management and stateful applications + +## Usage + +Load this skill when you need to: +- Understand Claude Opus 4.7's behavioral constraints and capabilities +- Reference safety guidelines and copyright rules +- Study tool usage patterns and routing decisions +- Compare system prompt designs across model versions +- Build agents with similar behavioral profiles + +## Notes + +This is a reference document extracted from the Claude Opus 4.7 system prompt. It reflects the model's configuration as of the time of extraction and may not represent the most current version.