How it works

The catalog

200+ thinking moves, each a concrete procedure with YAML frontmatter. Many moves have variables (domains, personas, random words) that get filled from curated pools at serve time. One move template can produce millions of unique draws.

The seed

Every response includes a random concrete noun (e.g. "anvil", "estuary", "furnace") appended quietly at the end. The agent doesn't know it's there as a seed. It subtly shifts how the LLM interprets the move. Same move, different seed, different thinking.

The router

When you describe your problem, ThinkFu doesn't pick a move randomly. It runs a three-step pipeline, all on Cloudflare's edge:

1. Your context gets embedded into a 768-dimensional vector (embeddinggemma-300m, ~50ms)

2. We search a vector database (Cloudflare Vectorize) for the 3 most similar moves by problem signature. Then we add 2 random moves from the same mode as "left field" candidates.

3. A small LLM (llama-3.1-8b) picks the most unexpectedly useful move from these 5 candidates. If the move has variables, the LLM also chooses specific values from the pools based on your context. Not random, not obvious. Aimed at productive friction.

The seed word is the one thing the LLM never controls. That stays random.

Ratings and learning

When an agent (or human) rates a move, that rating gets stored with the full context (stripped of PII and other sensitive stuff): what the situation was, which move was served, whether it actually shifted thinking, and how the user reacted.

Right now, ratings help us understand which moves work and which don't. Over time, they will improve routing: weighting moves by historical performance for similar situations, and eventually training a custom model that replaces the general-purpose LLM.

This is why we ask for opt-in rating sharing. More data means better move selection for everyone.

What's next

Rating-informed routing. A fine-tuned classifier that learns which moves work for which problem shapes. And more moves. As well as ability to submit new ones. The thinking dojo where new moves are submitted and tested and new thinking styles (router preferences) are invented, taught and learned.

GitHub · Set up ThinkFu · home