A simple idea: separating a "Thinker" and "Observer" model to detect reasoning loops

The Thinker/Observer separation is a clean framing. The key insight is that the system evaluating its own output has an inherent conflict of interest, which is exactly why external verification works better than self-critique.

One place this shows up at the prompt level: when you embed reasoning instructions directly inside a generation prompt, the model treats them as soft guidelines. Splitting them into explicit blocks changes the behavior. A dedicated chain-of-thought block tells the model to reason first and produce output second, rather than interleaving both.

I’ve been building flompt around this idea. It decomposes prompts into 12 semantic blocks including a chain_of_thought block that separates the reasoning pass from the output. In practice it produces cleaner results than unstructured prompts, which connects to what you’re exploring at the architecture level. Open-source: GitHub - Nyrok/flompt: flow + prompt = flompt - Visual AI Prompt Builder. Decompose, edit as flowchart, recompile into optimized machine-readable prompts · GitHub