Cut Your AI Token Costs by 10X With a Little Planning

Do you keep getting hit by rate limits on your tokens? Yesterday I walked through why a routine request like building a proposal deck can quietly burn through 163,000 tokens. Today’s about the fix: small, upfront decisions that can cut that number by 10x, on the exact same request.

I mentioned this in yesterday’s post too, but it’s worth repeating here because it’s the whole premise of today’s lesson: I run a system prompt in Claude that tells it not to produce any Microsoft file unless I specifically ask for it. Left to its own devices, it defaults to well-formatted Word docs before there’s any reason to build one — and that default costs real tokens every time. That one small rule is the same idea behind everything below, just applied more broadly: a little bit of upfront structure changes what the model has to do by default.

Same request, two paths

Both versions of the proposal-deck request start identically: same system prompt, same 150-token ask, same three inputs — a template, a call transcript, and meeting notes. From there, one path is a one-shot upload into a chat window. The other path takes a bit of prep work first. The gap between them is almost entirely about three decisions.

1. Stop feeding it PowerPoint

Any Microsoft file — a .pptx, a .docx — is essentially a zip file packed with formatting metadata, most of which has nothing to do with the content you actually care about. Feed that in every time you run the request, and the model has to read and reconstruct all of it, every single time. If you’re going to reuse the same template repeatedly, convert it once into something leaner — a JSON or XML version that captures the structure without the visual baggage. You pay that conversion cost once. You save on every run after that.

2. Excerpt instead of dump

The unplanned version pulls in the full 45-minute call transcript and the complete meeting notes, every word. The planned version calls a small tool that pulls only what’s relevant — the section of the call about pricing and scope, the lines in your notes about decisions. You lose some peripheral context doing this. You also stop paying to re-read forty minutes of small talk on every single pass.

3. Compile, don’t iterate

This is the one that surprised me most when I actually measured it. The unplanned version drafts the deck, then renders images to check its own formatting, almost always finds something wrong, fixes it, and renders again to confirm. That loop is expensive precisely because the model can’t see its own output without generating images of it. The planned version replaces the entire loop with one tool call to a compiler: it already knows the slide templates, it already knows what goes where, so it fills them in and produces the deck directly. No visual check needed, because the process guarantees the formatting is right. That single tool call costs next to nothing.

The number that matters

Same request, same deliverable: roughly 163,000 tokens without a process in place, roughly 16,000 tokens with one. That’s a 10x difference, and the planned version does take real setup time — building the template, the excerpt tool, the compiler. But that setup is a one-time cost. Once it exists, you get the savings on every proposal after that, which means at scale you can generate about ten proposals for the token cost it used to take to generate one.

Proposals are a good example because they sit at the top of the funnel — the faster you can turn one around, the better your odds generally look with a client. If your team is burning through a monthly limit on a single request, that’s not a sign AI doesn’t work for this. It’s a sign the process around the request hasn’t been built yet.

Key Takeaways

• Convert repeatedly-used files (PowerPoint, Word) into a leaner format like JSON or XML once, instead of feeding the raw file in on every run.
• Build a small tool that excerpts only the relevant sections of a transcript or notes, rather than pasting the whole document in every time.
• Replace draft-render-check-fix-render loops with a single compiler tool call that already knows the correct structure.
• The setup work is a one-time cost — the token savings apply to every future run of the same process.
• At a 10x reduction, you’re not just saving money — you’re able to do roughly ten times the volume within the same usage limits.

Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: Cut Your AI Token Costs by 10X