This repository holds the AI-client behavioural contract that the Draw Tree Protocol expects clients to follow. The same instructions ship in three wrappers so different runtimes can install whichever shape they understand:
| File | For |
|---|---|
SKILL.md |
Anthropic Skills format (Claude Code, Claude.ai, Claude Desktop, Perplexity Computer skills) |
AGENTS.md |
Codex CLI and any tool that follows the agents.md convention |
system_prompt.md |
ChatGPT custom GPTs, Claude project instructions, any "paste these instructions" surface |
The instruction body is identical across the three files. The wrappers only differ in:
- Front-matter (Anthropic Skills require
---YAML; AGENTS.md does not; a system prompt is just text). - A short trailer in
AGENTS.mdexplaining where Codex auto-loads it from.
The protocol spec (Draw-Tree/drawtree-protocol) defines what a tree must look like. This repo defines how an AI client behaves while building one — the entry gate, the strict ordering of Phase 1 stages, the never-chain-stages rule, the "never paraphrase the user" rule, etc.
You can change models (Grok 4, Claude Sonnet, GPT-5, Gemini), change AI clients (ChatGPT, Claude, Perplexity, Codex), or even change the server implementation, and as long as the client honours this skill, the research-design behaviour is the same.
mkdir -p ~/.claude/skills/drawtree
cp SKILL.md ~/.claude/skills/drawtree/SKILL.mdZip the file and upload through Settings → Skills → Upload skill.
mkdir -p ~/.codex
cp AGENTS.md ~/.codex/AGENTS.mdPaste the contents of system_prompt.md into your custom GPT's
instructions, your Project's instructions, or your client's system prompt.
This skill is written against the Draw Tree Protocol v0.3. It will be versioned alongside the protocol.
MIT.