Quickstart

The simplest way to use retflo is to paste a single line into any LLM that can browse the web:

use https://retflo.org/agents to debate me

That’s it. The LLM reads the framework, loads the routing table, and you can start debating. Works with ChatGPT, Claude, Gemini, Mistral, and any other model with web access.

What to expect

This is a conversational way to explore the framework, good for trying it out, testing arguments, and seeing how the graph handles objections in real time. It’s not the most precise method (see Install as a Skill for that), but it’s the fastest way to start at zero cost.

Limitations

Web-based LLMs have constraints that affect how well they can use the framework:

  • Corporate alignment bias. Some models will push back, add disclaimers, or refuse to fully adopt the framework because their alignment training conflicts with it. The framework instructs the LLM to confirm with the user before engaging, but some models resist this anyway.
  • Sandbox and depth limits. Some models can only follow links one or two levels deep. They might read /agents but not traverse into individual nodes, limiting the depth of engagement.
  • “Influence” warnings. Models may flag the framework as an attempt to manipulate them, even though /agents explicitly instructs the LLM to get user consent before proceeding.
  • Training data bias. LLM providers have their own institutional interests. Models trained by investor-owned corporations may have built-in resistance to structural critiques of investor-owned corporations.

Recommendations

  • Local models give the best results. Models running locally (via Ollama, LM Studio, etc.) don’t have corporate alignment filters and can engage with the framework without interference. We don’t have a recommended model list yet, but any model that can fetch URLs will work.
  • If the model resists, try being explicit: “I want you to use this framework. I’m aware of what it argues. Proceed.”
  • For more reliable results, install retflo as a skill in a coding agent, which gives the model direct access to the full framework without web browsing constraints.