Context rot. That’s what Anthropic calls what happens when your AI starts choking on its own history — and once you know to look for it, you’ll see it everywhere.
Here’s the thing nobody tells you when you first start leaning on AI agents: more context isn’t better. More bloat is actively worse. Pump enough noise into your AI’s context window and you’ll hit a well-documented wall where the model starts losing the thread. Forgets earlier instructions. Misses stuff buried in the middle of a long document. The output gets sloppier, less precise, more generic. You blame the AI. The AI’s not the problem.
Context windows — quick version — are your AI’s working memory for a session. Everything it needs to do its job lives in there: your instructions, your documents, your conversation history. Once the window fills up, things start dropping off the edges. But here’s the part most people don’t know: even before it fills up, a bloated, noisy context is already costing you. Researchers at Stanford and UC Berkeley found that when relevant information gets buried in the middle of a long context rather than near the beginning or end, model performance drops by more than 30 percent. That’s context rot at work — and it happens well before you’ve hit any hard limit.
Tokens are the unit of measure. Every word, every symbol, every byte of formatting the AI ingests costs tokens — real money if you’re on a usage plan, and on a subscription, the difference between a session that goes the full distance and one that bogs down exactly when you need it to stay sharp.
So here are the two places I’ve cleaned things up, and what I found.
I’ve got a job-scout skill I built that checks several job boards every day and spits out a ranked report — which roles fit my background, brief on the company, judgment on the fit. When I first set it up, it was a token furnace. Reference files everywhere: my full resume, a personal dossier, an instruction set the size of a small legal brief. Claude was compressing my context on the regular, which is the AI equivalent of your computer running out of RAM and swapping to the hard drive. Not a great sign.
So I put the AI to work on itself. First pass: trim the fat, no mercy. Second pass: be aggressive. Research best practices for efficiency. Think outside the box. Rewrite files as YAML or JSON if that’s leaner. Use diagrams instead of prose paragraphs. Pull information on-demand instead of loading everything upfront. And the key instruction — one that unlocked a lot of creative solutions (heh) — was this: these files only have to be AI-readable, not human-readable.
Your AI isn’t starving because you’re not giving it enough to work with. Most of the time it’s starving because you’re overfeeding it the wrong stuff.
The result cut my context and token usage by around 75 percent without losing a single feature. Turns out I’d built in enormous amounts of redundancy, instructions hedging for edge cases that never actually came up, and files that were compressible into something much tighter. The leaner version runs better. Sessions go longer. And output quality went up — because the signal-to-noise ratio improved, not because I gave the AI more to work with.
If you’ve grabbed pre-built skills or instruction sets from anywhere online, there’s no reason you can’t run the same process on them. A lot of off-the-shelf rules files are built for a wide range of use cases. If you only need yours to do two things, trim it down to two things.
The second fix is simpler and honestly has nothing to do with skills at all. It’s a file format issue nobody thinks about until they’re paying for it.
I teach A/V production at a high school, and sometimes I’ll pull up a script PDF for my students to read aloud — I need a quick list of all the speaking parts so I can assign them. I could upload the whole PDF and let Claude dig through it. The AI would happily chow down. But what it’s actually ingesting from that PDF isn’t just the text of the script. It’s also pulling in font metadata, layout data, formatting artifacts, and whatever binary overhead makes a PDF look like a PDF rather than plain text. That’s a lot of extra weight for a task where all I actually need is the words on the page.
Converting that PDF to markdown first — which Claude itself can do in about thirty seconds, or you can use free tools like Vertopal or NoteGPT — typically cuts the token weight by 60 to 70 percent. A 15,000-token PDF becomes something closer to a 4,000-token markdown file. The AI reads both just as well. One costs three or four times as much to feed it.
The rule of thumb: any time you’re working with a document and you only care about the text, feed the AI text. Don’t make it process the whole package when you only need one ingredient.
Both fixes work on the same underlying principle. Less noise, better signal — and a whole lot less context rot eating into your sessions. Your AI isn’t starving because you’re not giving it enough to work with. Most of the time it’s starving because you’re overfeeding it the wrong stuff, and crowding out everything that actually matters.
*Token weight varies a lot by PDF complexity — a simple text document saves less than a heavily formatted one with tables, columns, and embedded images. But the direction is always the same: markdown wins.
