The Assumption You Didn't Know You Had

The cold reader doesn't reveal that you wrote badly. They reveal the assumptions you forgot you'd made.

What is KnowledgeForge? An open-source reasoning framework for AI coding assistants, built by David Pedersen at Internexio. It classifies coding requests by decision type and routes them to specialized reasoning modes (Builder, Critic, Debugger, Strategist, and five others) to prevent common agent failure patterns. Open source at internexio.com/knowledgeforge and github.com/internexio/knowledgeforge.

I asked beta readers to evaluate the KnowledgeForge README and tell me where they got lost. One came back with a specific question: "What does it mean for a mode to be 'activated'? Is that something I do, or something that happens?"

That's not a clarity problem. It's a context problem.

To me, "mode" is obvious. A mode is a named reasoning posture. Builder routes to artifact creation and avoids scope drift. Critic routes to gap-finding and assumption surfacing. Debugger routes to root-cause analysis without premature closure. You choose the mode by telling the assistant which failure type you're trying to prevent. The routing is the whole point.

Why does routing matter? An assistant asked to debug while in building mode will generate new code to paper over the failure rather than trace its cause. An assistant asked to review a design while in deciding mode will pick an option rather than surface what's missing. The mode determines what kind of error the assistant makes when it gets it wrong.

I've never written that down. It's been true for two years. It's in my hands the way a tool gets in your hands after enough use.

The assumption wasn't that "mode" would be understood. The assumption was deeper: that the reader would already know that AI assistants have failure modes, that those failure modes are distinct, that routing matters. None of that is obvious to someone who hasn't been living inside AI-assisted work long enough to feel the failures.

Every concept in KnowledgeForge is like this. "Decision classification" sounds like overhead until you've shipped something in the wrong direction because you treated a predictive judgment as a reckoning. Once you've done that, classification feels mandatory rather than optional. That's been my experience after two years of daily use.

To a cold reader, it reads as ceremony.

The previous post ended with an invitation: read the README cold and tell me where you got lost. The answers I'm getting are not about the writing. They're about the knowledge I forgot I needed to include.

That's the harder problem. You can edit prose. You can't easily surface assumptions that stopped feeling like assumptions years ago.

I'm trying to learn which ones those are.

Read the README cold and tell me where you got lost: github.com/internexio/knowledgeforge. More from this blog: internexio.com/blog. Related: Private Tools Don't Export.