“When you’re asking it a question, it’s not looking for the answer, it’s looking for the process. And it’ll bring that and run the process the way that you want it to. Now you could have a lower quality model, because all the reasoning’s embedded into your markdown files.” — Lou
Session context: 2026-07-23_Mastermind — Jay was describing a RAG-plus-local-LLM architecture he was designing for writing, and Lou followed the thread back to how DeepSeek trained its reasoning model. The idea arrived live, mid-sentence, and Lou spent the rest of the session circling back to it.
Core Idea
Everyone building a knowledge base is optimizing the wrong retrieval target. You embed your documents, you tune your chunking, you argue about semantic versus keyword — all so the system can find the answer faster. But an answer is a dead artifact. It was correct for one question, at one moment, under one set of assumptions, and it tells the model nothing about what to do when the next question is slightly different.
The reframe came from an observation about how reasoning models actually got built. When DeepSeek distilled a reasoning model out of frontier models, the training data was not the outputs. “They were not asking Claude to produce an answer, they were asking Claude to produce the reasoning as output. And that’s what their training data was.” Reasoning traces, not conclusions. That is also why the frontier labs stopped letting you download the chain of thought — the trace was the valuable thing, and it was walking out the door.
Now apply that to your own stack. If you keep a skills library — markdown files that describe how you do things — you already have a corpus of reasoning traces. They just aren’t wired up as a retrieval target. Make the skills library the knowledge base itself: ask a question, and instead of returning a passage that contains an answer, the system returns the procedure for producing one. The model then executes the procedure.
The consequence is the part worth sitting with. Whatever reasoning capability the model lacks, you can supply from the file. A cheaper, dumber, faster model that retrieves a good process and follows it will outperform an expensive model improvising from scratch — because the expensive model’s advantage was mostly in constructing the process, and you already paid for that once. The routing table replaces the vector store. There is nothing to embed; there is a map from problem shapes to procedures.
This also sidesteps the fine-tuning trap entirely. Fine-tuning bakes a style into weights you cannot inspect and cannot revise without retraining. A process file is legible, editable, and version-controlled — and it upgrades the moment you learn something better.
Practical Application
Take one question a client or colleague asks you repeatedly. Don’t write the answer. Write a markdown file called how-to-answer-<question>.md containing only the steps you take to reach an answer — what you check first, what disqualifies an option, what you refuse to conclude without evidence, what the output should look like. Then hand that file plus the question to the cheapest model you have access to (Haiku, Sonnet-low) and compare its output to what your most expensive model produces cold. If the cheap-model-plus-process wins, you have just found the actual location of your intelligence — and it wasn’t in the model.
Related Insights
- Insight - Ambient Knowledge — Why Ambient Intelligence Is Judgment Transfer, Not Retrieval — established that ambient intelligence transfers judgment rather than retrieving facts; this names the retrieval mechanism that makes it operational.
- Insight - Prompt the Cheap Model With the Smart Model — Pay Once for Reasoning, Reuse It Forever — the same economics, one level up: there the smart model writes the prompt, here it writes the process file.
- Insight - RAG Is Raw Material, Not Answers — Design for the Right Retrieval Architecture — argued RAG returns raw material for inference; this argues the raw material should be procedure, not prose.
- Insight - The Markdown Test — If It Can’t Be Written Down, It Belongs to the Harness — the portability test that makes process files worth writing at all.
- Insight - Fine-Tune for Voice, RAG for Capability — the companion claim from the same session: fine-tuning moves style, retrieval moves capability.
- Insight - The More Inference You Pull Into the Model, the More Model Churn Owns Your Product — the risk this insight mitigates: reasoning in files is reasoning that survives a model swap.
Evolution Across Sessions
Builds on Insight - Ambient Knowledge — Why Ambient Intelligence Is Judgment Transfer, Not Retrieval (2026-04-02), which established that the value of a structured knowledge base is judgment transfer, and on Insight - Prompt the Cheap Model With the Smart Model — Pay Once for Reasoning, Reuse It Forever (2026-06-11), which established that reasoning is a one-time purchase you can amortize. The new development is the retrieval target itself: make the skills library the OKF, so the query resolves to a process rather than a passage, and the model’s own reasoning ceiling stops being the binding constraint. Lou explicitly flagged this as an untested hypothesis — “I’m kind of hypothesizing this” — and said he intends to experiment. Future sessions should report whether cheap-model-plus-process actually beats expensive-model-cold on real work, and where the approach breaks down (Jay’s caution: orchestration-level judgment may not survive the downgrade).