Your AI Re-Reads Your Whole Setup Before Every Answer | Edition 329

You spent twenty minutes writing the perfect custom instructions. You built a Project and fed it every reference document that matters. You turned on connectors for your calendar, your notes app, your email — because why wouldn't you, they were right there. And then the assistant started giving you answers that felt off. Vaguer. Slower to get to the actual point. Less sharp than it was on day one, back when you'd barely told it anything at all.

That is not your imagination, and it is not a coincidence. It is a direct, mechanical consequence of what you built — and understanding why is the fastest way to fix it, because the fix is not a better prompt. It is subtraction.

Why "more attached" does not mean "more capable"

Start with the part that is easiest to miss: the model itself has no memory between messages. OpenAI's own developer documentation says it plainly — "each text generation request is independent and stateless." Anthropic's engineering team describes the same reality from the builder's side: assembling what the model sees is iterative, and "the curation phase happens each time we decide what to pass to the model." Whatever feels like an ongoing conversation is an illusion built by an app quietly resending everything relevant — your custom instructions, your history, and, for smaller Projects, your reference files — with every single call.

That mechanism is true across chat assistants generally. What differs is how much of "everything you attached" a resend actually includes, and that is where the two most popular assistants genuinely diverge — not a place a summary should flatten together.

Anthropic's own help documentation for Claude's Projects feature says a small Project gets read in full, but "when your project knowledge approaches the context window limit, Claude will automatically enable RAG mode" — retrieval, not a full re-read — without publishing the exact token threshold where that switch happens. ChatGPT's Custom GPT "Knowledge" files work differently from the start: OpenAI's help center describes them as chunked and turned into embeddings for retrieval the moment you upload them, regardless of size, so "the GPT can access the uploaded files to get additional context" by pulling only the relevant pieces rather than the whole library. Two real products, two different answers to "does it reload the whole thing" — worth knowing rather than assuming either way.

What does not get that same protection, in either product, is your custom instructions and whatever apps you have connected. Those are short enough that nobody builds a retrieval system around them — they get resent whole, every time. Anthropic's own engineering write-up on tool use says as much about connected tools specifically: "most… clients load all tool definitions upfront directly into context" before the model ever sees your question, and in the extreme case of an agent wired to thousands of tools, "they'll need to process hundreds of thousands of tokens before reading a request." Almost nobody is running thousands of connectors — but the mechanism is identical at any scale: every connected app adds its name, its description, and its full parameter list to what gets loaded, whether or not you touch it that day.

Put those pieces together — instructions and connectors reloaded whole every time, plus a knowledge base that may or may not get the retrieval treatment — and you get a real number, even if you have never seen a token counter. The context window is finite. As of this writing, Anthropic's current flagship Claude models ship a 1 million‑token window (its smaller Haiku model still runs 200,000); OpenAI's current flagship line publishes roughly 1.05 million. Those figures sound enormous, and for a single document they usually are. They stop feeling enormous once you are stacking a long instruction block, a Project's worth of files, and a dozen connector descriptions before you have typed a single word — because whatever slice that stack consumes is not "used later." It is already spent before your actual question gets read.

Here is the part that makes this worth fixing rather than just noting: it is not only about running out of room. Anthropic's own engineers, in a 2025 write-up on building agents, describe something they call context rot — "as the number of tokens in the context window increases, the model's ability to accurately recall information from that context decreases," a pattern they say "emerges across all models," even though "some models exhibit more gentle degradation than others." That is not an outside critique — it is Anthropic saying it about its own product category. Independent researchers at Chroma tested the claim directly: across 18 models and several realistic long-context tasks, they found performance consistently, if unevenly, degraded as input length grew, even holding the actual difficulty of the task constant. It is worth being precise about what this does and does not say — it is a gradient, not a cliff, and it shows up on benchmark-style recall and reasoning tasks rather than every kind of everyday question. But the direction is real, and it is not folklore: bulk in the context window is a cost against quality, not just a rounding error against speed.

So when a heavily-configured assistant starts feeling vaguer than a blank one, that is not a worse model answering you. It is the same model, working harder to find your actual question inside everything you buried it under.

Below is a rough, illustrative model of how that stacks up. Toggle a few realistic setups and watch how much of the window is spoken for before you have asked anything. It is a model built to make the shape of the problem visible, not a live measurement of any specific account — treat the numbers as illustrative, not as something anyone actually measured on your setup.

The ten-minute audit

You do not need a token counter to check any of this yourself. Everything below is something you can look at directly, on your own account, in about ten minutes — and each one tells you something different, so do not skip to the end.

1. Your custom instructions (or "personalization" / "custom instructions" settings). What to look for: open the actual settings screen and read the whole thing end to end, not a skim. What too much looks like: it has grown into several paragraphs covering situations that do not apply to most of what you ask, or you cannot summarize it in one sentence without needing an "and also."
2. Your Project's file list (or a Custom GPT's Knowledge files). What to look for: count the documents, then note which ones you have actually opened, referenced, or been asked about in the last month versus which just sit there. What too much looks like: more than a couple of files nobody has touched recently, or documents kept "for completeness" that do not relate to what you actually use this Project for.
3. Your connected apps, switched on. What to look for: the full authorized list in your settings, not just what you remember turning on. What too much looks like: anything you enabled once for a single task and never disabled, or a name on the list you would not recognize if asked what it does.
4. Which of those you have actually used this month. What to look for: cross the "switched on" list against what you have genuinely relied on in a real conversation recently. What too much looks like: any connector that is switched on but not on the "used" list — that is pure overhead with no offsetting benefit, paid every single time.

If items 3 and 4 do not roughly match, or if you cannot get through item 1 in under a minute of reading, you are carrying weight that is not earning its keep.

None of this means custom instructions, Projects, or connected apps are mistakes. They are the whole reason these tools became useful enough to build a routine around. It means they are not free, and the bill does not show up as a slower response. It shows up as a duller one. The fix is not cleverer prompting. It is periodically checking what is actually in there, and cutting whatever is not earning its seat.