Building in Public: Why Our First Attempt at Automating Thinking Patterns Failed
At NAKPRC, we believe in building in public and sharing our hard-learned lessons. Here is the story of how our initial attempt to automate the generation of our Thinking Patterns SDK using Qwen 3.6 and Claude Code resulted in a total research failure.
At NAKPRC, our philosophy is Open Knowledge and Building in Public. This means we don't just share our successful launches; we openly document our experimental failures. Today, I want to talk about a major research pivot we had to make while building the Thinking Patterns SDK.
The Original Hypothesis
When we conceptualized Thinking Patterns—a strategic AI knowledge platform and modular ecosystem—our goal was ambitious. We wanted to build a robust suite of npm packages (llm-thinking-patterns-nakprc and nakprc-thinking-patterns) that developers could use to seamlessly integrate complex cognitive frameworks into their applications.
Our initial hypothesis was that we could fully automate the logic structuring of these frameworks. By utilizing automated code-generation loops powered by cutting-edge models like Qwen 3.6, Ollama, and Claude Code, we believed we could rapidly generate the core SDK without tedious manual intervention.
In theory, feeding a complex cognitive schema into the AI would yield perfectly structured, deterministic JavaScript logic.
The Research Failure
In practice, the 1st attempt failed completely.
While AI code generation is phenomenal for scaffolding standard boilerplate or single-file scripts, it struggled immensely with the depth of deterministic logic required for our SDK.
When pushed to map edge-case reasoning pathways across a complex taxonomy, the automated output yielded inconsistent structures. Worse yet, the models frequently hallucinated non-existent logic pathways or drifted from the required data schemas. An SDK needs absolute reliability, and the automated AI approach broke that core tenet.
We had to acknowledge the limitations of current automated generation for complex reasoning schemas. This was a research failure.
The Pivot: Manual Engineering
Failure in engineering isn't a dead end; it's a redirection.
Acknowledging that the models couldn't reliably build the underlying logic engine, we decided to pivot. We abandoned the automated code-generation approach and have transitioned entirely to a manual, deterministic engineering process.
This means writing, validating, and testing every single edge-case reasoning pathway by hand. It's slower, harder, and less flashy than claiming an AI built our SDK for us—but it's the only way to ensure the absolute fidelity and reliability that developers expect from NAKPRC.
The Thinking Patterns project is far from abandoned. In fact, we are more committed to it than ever, and we will continue working vigorously to launch these packages to the community.
Quality software is measured in years, not weeks. Sometimes, you have to do things the hard way.
1st Attempt RESEARCH Failure using AI (ollama's claude code using qwen3.6 within 1day)
