I have a confession that will surprise no one who has met me: I let my AI re-read my entire business every time I asked it a question.
Not on purpose. It just happened. Sixty blog drafts. The business plan, the marketing plan, the brutal self-assessment I commissioned and then had to live with. Contracts. Client notes. Pricing tiers. Every time I asked Claude something like “what would I quote this prospect, and what’s our reusable pattern for their problem,” it would go read the relevant pile of source material, reason its way to an answer, hand it over — and then forget all of it. Next question, same archaeology. The model was rediscovering my own company from scratch, on my dime, several times a day.
This is the dirty secret of “just put it all in context.” It works, and it’s also the most expensive possible way to be right.
So I built it a wiki. Not a wiki for me — a wiki for it. The pattern’s been making the rounds; Karpathy wrote it up. The idea is dead simple: instead of letting the model retrieve raw documents at query time, you have it compile everything once into an interlinked set of pages it maintains itself. New source comes in, the model files it — updates the summaries, flags the contradictions, fixes the cross-references. The bookkeeping that makes every human wiki die of neglect is exactly the chore an LLM does tirelessly and for roughly nothing.
The payoff is real and it’s large. Take that quoting question. Answering it from the wiki means reading the index and about five small compiled pages — roughly 1,530 tokens. Answering it by retrieving raw chunks the RAG way runs about 8,000. Answering it the way I’d been doing it — load the corpus, let Claude sort it out — is about 185,000 tokens. Same answer. One hundred and twenty times the cost.
Here’s where the story would normally end, with me looking clever. Except I’m constitutionally incapable of leaving a victory lap un-sabotaged, so I plotted the cumulative cost — and the wiki immediately got off its high horse. Because the wiki wasn’t free. Compiling the whole corpus the first time cost about 220,000 tokens. One giant reading assignment, paid up front. Against the “dump everything every time” approach, that pays for itself by roughly the second question. But against a competent RAG setup that only spends 8,000 a query? The wiki doesn’t pull ahead until somewhere around the thirty-fourth question.
So the honest headline isn’t “I made my AI 120× cheaper.” It’s “I spent 220,000 tokens building a thing to save tokens, and it won’t break even against the obvious alternative until I’ve used it three dozen times.” Which I will, easily — but I’m not going to pretend the upfront bill didn’t happen.
If it were only about tokens, RAG would be the lazy, correct answer and this would be a cautionary tale. But the token count quietly understates the wiki, because it assumes the two approaches return the same answer. They don’t. RAG hands the model twenty disconnected fragments and asks it to reconstruct the big picture every single time — re-deriving which pricing tier applies, re-noticing that two documents contradict each other, re-discovering that this client’s problem maps to that reusable pattern. The wiki hands back an answer where all of that work is already done: synthesized, cross-linked, cited, reconciled. The contradictions were flagged during the compile, once, instead of rediscovered during every query, forever.
The thing I keep relearning in 2026 is that the expensive part of working with AI is almost never the asking. It’s the context — assembling it, paying to re-send it, and watching the model burn reasoning to rebuild understanding it already had an hour ago. Compiling that understanding once and keeping it current is the whole game.
So yes: I optimized my AI’s reading homework by first making it do an enormous reading assignment. I’m aware of the shape of that. But the corpus only gets read cold once now, and every question after is a model that already knows my business instead of one frantically re-learning it.
I taught it to remember so I’d stop paying it to forget. Worth every one of those 220,000 tokens — a sentence I’ll be repeating to myself, for reassurance, around query thirty-three.
Numbers are order-of-magnitude, not billing-exact: token ≈ characters ÷ 4, measured from the actual files; RAG modeled at ~20 chunks × ~400 tokens; the build cost is the real full-corpus ingest. Charts are static inline SVG — no scripts, no external dependencies.