A 14-minute explainer from Ben AI (Ben van Sprundel) on the single habit that most improves AI output quality: managing context rot. The core argument is that Claude gets measurably worse the longer a chat runs, well before the context window is full, and the fix is to track your token usage and start fresh chats without losing your accumulated context.
The problem: context rot
Most people keep a few giant chat sessions they return to for repetitive work, because starting fresh means re-explaining all the rules and context. The hidden cost is context rot. Every prompt, answer, file, connector, and skill fills the context window, and the fuller it gets, the more the model forgets, blurs, and drifts. Ben's analogy: the context window is like your brain over a day. The more you cram in, the fuzzier and less productive it gets. Opus 4.8 has a 1 million token window, but quality drops far before that limit.
His rough zones (rules of thumb, not official numbers):
| Tokens used | Zone | Output quality |
|---|---|---|
| 0 to 200k | Sweet spot | Best outputs |
| 200k to 350k | Usable | Starts drifting |
| 350k+ | Degraded | Noticeably worse |
| 400k+ | "You're right to push back" | Start a fresh chat |
The tell: when Claude keeps agreeing that you're right to push back, you are probably past 400k and in the "dumb zone."
Know where you are: the /context habit
You cannot manage what you cannot see. In the terminal Claude Code shows token usage, but Claude Cowork and the desktop UI do not, by default. The fix is the /context command, which works in any chat in Cowork or Claude Code and reports how many tokens you have spent and on what.
Ben's advice is to build the habit of running /context regularly. Beyond telling you when to start fresh, it teaches you what actually burns tokens. His own lesson: research, connectors, and MCP calls are among the biggest spenders, so he switches chats much sooner when using them.
Four signals to run /context:
- Long sessions (the obvious one).
- Recurring chats you keep returning to because they are already primed with your rules (his newsletter-review chat, his per-video ideation chat).
- Heavy connector or research use, which burns tokens faster than you think.
- Outputs visibly getting worse, having to re-explain, or you getting frustrated. This is the strongest signal, and the moment to check rather than to keep arguing.
Why /compact falls short
Claude's built-in /compact summarises the chat and context into roughly 50 to 60k tokens, then continues in a fresh context window behind the scenes. Useful in theory, but Ben lists three flaws that made him avoid it and just tolerate rot in long sessions:
- The summary is too thin, so you end up re-explaining much of the same context anyway.
- The little rules, do's, and don'ts you accumulated across the back-and-forth get lost.
- Context docs you fed in get reduced to tiny summaries, so you have to re-feed them.
He also notes /compact only runs in Claude Code, not Cowork, where he does most of his knowledge work.
The Re-fresh skill
Ben built a free Re-fresh skill to fix those gaps. It is the same idea as /compact, a summarise-and-continue handoff, but built for knowledge work:
- Works in Cowork (and Claude Code), not just the terminal.
- Writes a fuller summary (around 100k tokens). It costs more, but it stops you re-prompting the same context.
- Asks the goal of your next session first, then saves the context relevant to that goal.
- Captures the accumulated rules: where the task stands, what worked, what did not, corrections made, and the specific do's and don'ts you gave.
- Lists every context doc used and instructs the next session to read them all first, instead of leaving you to re-feed them.
In practice it drops a set of summary files into your working folder (outline, research, decisions, rules) and generates a ready-to-paste prompt for the new chat that states the goal, tells Claude to read those files first, and points to your other context (strategy, ICP, brand, transcripts). Paste it into a fresh chat with the same folder selected, and the new session is primed to continue.
Two ways to use it
- Same task, fresh chat. When a long session hits the dumb zone, hand it off and keep going without the rot.
- New task, same context. When you want to veer off, say to write a newsletter from a YouTube-ideation chat, do not pollute the original chat. Refresh into a new window that keeps the context but starts clean for the new task.
Key takeaways
- Rot starts early. Quality degrades from ~200k tokens and gets bad past ~400k, long before the 1M window fills.
- Make
/contexta reflex. It shows where you are and trains your intuition for what burns tokens. - Frustration is a signal, not a bug. When outputs slip or Claude keeps conceding, switch chats instead of arguing.
- A rich handoff beats a thin summary. Carrying the rules and the actual context docs, not a paragraph recap, is what makes a fresh chat as good as the primed one.
- Always work inside a selected folder in Cowork or Claude Code, so a handoff has somewhere to write its files.
Notable quotes
"Essentially, it becomes dumber the more you use it, which you've probably noticed."
"Instead of screaming at AI, you have to get in the habit of using /context."
Why it matters for me
I live in long Claude Code and Cowork sessions building skills, notes, and research, exactly the recurring-primed-chat pattern Ben describes, so this is directly useful. Two changes worth adopting: run /context as a habit to catch rot before it wrecks an output, and treat a folder-based handoff as the default way to continue or pivot a task rather than dragging a bloated chat along. Credit where due: Ben says the Re-fresh skill was inspired by Matt Pocock's handoff skill, which does the same for engineering work and is worth a look for coding sessions.
Further Reading
