Skip to content

generic.md

Latest commit

 

History

History
45 lines (33 loc) · 2.55 KB

generic.md

File metadata and controls

45 lines (33 loc) · 2.55 KB

Install — Generic (Any Agent/Framework)

Recommended: Run ./install.sh from the repo root for automated installation. Supports Claude Code, Gemini CLI, Codex CLI, Cursor, OpenCode, and OpenClaw. The manual steps below are for reference or troubleshooting.

Universal principle

Double-Loop Learning is a cognitive base — it changes how the agent reflects, not what tools it uses. Installation means injecting cognitive-protocol.md into the agent's always-on instructions (system prompt, rules file, or configuration).

Platform mapping

Platform Where to inject File to use
Claude Code ~/.claude/double-loop-learning.md + ref in CLAUDE.md cognitive-protocol.md
Codex Prepend to AGENTS.md cognitive-protocol.md
Gemini system_instruction field cognitive-protocol.md
Cursor Prepend to .cursorrules cognitive-protocol.md
ChatGPT Custom Instructions -> system prompt cognitive-protocol.md
LangChain System message in chain cognitive-protocol.md
AutoGPT / CrewAI Agent system prompt cognitive-protocol.md
Any other Find the "always-on instructions" config and inject there cognitive-protocol.md

Step by step

  1. Locate your agent's system prompt or always-on rules file
  2. Copy the contents of cognitive-protocol.md (~30 lines)
  3. Paste it into the system prompt, BEFORE any domain-specific instructions
  4. If using other cognitive bases (First Principles, Systems Thinking), place them first, then Double-Loop Learning

File selection guide

Need File Size
Minimal install (core rules only) cognitive-protocol.md ~30 lines
Full reference framework + SKILL.md ~200 lines
Anti-pattern detection + anti-patterns.md ~150 lines
Teaching examples + examples.md ~150 lines

For most agents, cognitive-protocol.md alone is sufficient. The additional files are reference material for when the agent needs deeper guidance.

Troubleshooting

  • Agent ignores the rules: Move cognitive-protocol.md content to the beginning of the system prompt, not the end. Most models weight earlier instructions more heavily.
  • Rules conflict with domain instructions: Double-Loop Learning should never conflict — it changes reflection process, not output format. If conflict appears, the domain instruction likely contains an unexamined assumption. That's a feature, not a bug.
  • Context window pressure: cognitive-protocol.md is ~30 lines. If that's too much, something else in your prompt needs trimming first.