Type: regular Meta Title: Why I Restart AI Sessions Every 90 Minutes for Better Output
I have been using AI coding and writing tools heavily for the past 18 months. The pattern I noticed roughly nine months in was that my sessions started strong and got progressively worse the longer they ran. The first hour produced sharp, useful output. The second hour produced output that was technically correct but slightly off. The third hour produced output that was clearly degraded, even though the model was unchanged. The issue is not the model. It is the context window.
Every message in a session adds to the cumulative context the model is paying attention to. The tool reads previous messages, previous tool calls, file contents, and its own prior responses all together when generating the next reply. Past 60 to 90 minutes of active work, the context window starts to fill with the model's own thinking from earlier in the session. The attention available for your current question gets diluted across all that history. The output quality drops without producing any visible error.
The lost-in-the-middle problem is the other half. Information placed early in a long session migrates from the salient position at the start of the context to the less attended middle. Architecture decisions I made in message 5 of a long session get treated like background noise by message 80. The model still has the information. It cannot use it as reliably. Decisions get reversed or contradicted in subtle ways.
The fix is mechanical. Start a new session every 90 minutes. Lead the new session with a tight context-setting message that summarizes the architecture, the current task, and the relevant files. Do not paste full files unless necessary. The cost of the restart feels real because the conversation history goes away. The benefit is that the next 90 minutes produce work as sharp as the first hour did.
I run this in 90-minute blocks now. Block 1 from 8 to 9:30 AM. Restart. Block 2 from 9:30 to 11 AM. Restart. Block 3 from 11:30 AM to 1 PM. Lunch. Block 4 from 2 to 3:30 PM. Restart. The total output across four blocks is dramatically higher than what I was producing in two long sessions before, even though the total clock time is the same. The mechanism is that each block actually runs at peak quality instead of degrading through the second half.
The other practice that compounds is externalizing state. Architecture decisions, naming conventions, and project context live in a markdown file the AI reads at the start of each session. The file is one to three pages. The model loads it as fresh context. The same information that would have lived in conversation history now lives in a stable file. Resets become much cheaper because the file carries the context forward.
For developers and operators using AI coding tools heavily, the productivity gain from this discipline is real and measurable. I tracked my output for two weeks under the 90-minute reset protocol and compared it with my prior pattern of long open-ended sessions. Ship-rate on completed tasks went up about 35 percent. Quality on the work that did ship was meaningfully better. The work felt less exhausting at the end of the day, because the mental load of maintaining a long session never had a chance to accumulate.
For Christian operators thinking through how to use AI tools responsibly, the 90-minute discipline also pairs well with the human work that surrounds the AI. Reviewing AI output critically requires attention. Attention is finite. The 90-minute reset protects the attention you need to actually evaluate what the model is producing rather than rubber-stamping whatever it generates.
The tools were built for sprints. Treating them as marathons produces worse work. The reset is the highest-return change you can make to your AI workflow this quarter. Start tomorrow. Set a 90-minute timer. Restart when it rings.
