Private Tools Don't Export

KnowledgeForge is an open-source reasoning framework for AI coding assistants, built by David Pedersen (SEMalytics). It routes each task to a specialized reasoning mode, preventing the most common AI agent failure: mis-routing the type of problem the model is solving. Repo: github.com/internexio/knowledgeforge.

Someone asked me to summarize it in an email. My reply was better than the README. That’s the problem this post is about. KnowledgeForge is a reasoning harness for AI coding assistants. It started in 2023 as a stack of instruction files I kept for myself: patterns I noticed, things I wanted the model to carry between sessions. By late 2025 it had 43 modules.

Forty-three modules is not a framework. It’s a documentation project.

I ran a Pareto pass in late 2025: which modules produced measurable results in daily work, which were overhead for edge cases that never happened twice. Nine survived. The other 34 were removed because they covered failure modes that appeared once across months of production use and never recurred. Coverage anxiety, not recurring problems.

Nine modes, each targeting a different type of reasoning failure:

• Builder — creates artifacts (code, specs, implementations); starts from the problem, prevents over-engineering.
• Critic — reviews work for gaps, contradictions, unstated assumptions; read-only, does not modify.
• Debugger — root-cause analysis through hypothesis testing; requires 0.8+ confidence before declaring a cause.
• Strategist — evaluates options with explicit trade-offs; forces prioritization over “do everything.”
• Expert — adversarial depth analysis for irreversible or high-stakes decisions.
• Synthesizer — extracts reusable patterns from multiple examples; every pattern ships with anti-patterns.
• Calibrator — configures AI tools with guardrails right-sized to project complexity.
• Coordinator — designs multi-agent workflows by mapping dependencies first.
• Navigator — resolves genuinely ambiguous requests with one targeted question.

Each mode routes differently because the failure it prevents is different. Debugging routes to hypothesis testing. Reviewing routes to adversarial gap-finding. Deciding between options routes to forced trade-off analysis. The compression was real.

The README wasn’t.

I open-sourced KnowledgeForge in early 2026. First real test: write a README for someone who doesn’t already know what it is. A reader who couldn’t parse the README cold asked me to summarize it in an email. I wrote three paragraphs. They were better than anything in the file.

That’s the thing with tools you build for yourself. They’re calibrated to the context in your head. Going public doesn’t mean publishing the files. It means explaining what you already know.

The best explanation I’ve written of KnowledgeForge is that email. Not the README. I rewrote it today.

Repo: github.com/internexio/knowledgeforge. Read it cold and tell me exactly where you got lost. That’s the test I’m still trying to pass.

Related reading on this blog: The Maintenance Bet and KnowledgeForge Open Source: One Week In. Also see the KnowledgeForge project page.