As a product manager, you’re probably very familiar with the hazards of context switching. A good engineering leader will coach you to protect engineers from too much context switching, but you’ll likely have personally felt the effects yourself, too.
Our roles mean we sit at the intersection of every function – engineering, design, marketing, sales, customer support, leadership. Which often means our days are built from fragments: a bug triage, a roadmap review, a customer call, a stakeholder update. Each of the back-to-back meetings comes with an abrupt change in context, and that’s before we consider the firehose of messages in between.
With that constant flow of input throughout the day, our brains have only a limited capacity to meaningfully hold on to the information we gather, and to make effective judgements based on it. That is a particularly expensive problem, because product work is so judgment heavy. Whether it’s synthesising, prioritising or making trade-offs, judgement is critical.
There’s an echo here in the world of AI. We know that large language models can only process a finite amount of information at one time. That fixed “context window” gets consumed by the length of a prompt, its accumulated conversation history, any content retrieved from referenced sources, and the model’s response. As the window approaches capacity, “context rot” sets in.
Interestingly, the “lost in the middle” effect means that information stored at the beginning or end of the input context suffers less from context rot. But performance significantly degrades when models need to access information from the middle of long contexts.
AI researchers have identified strategies to limit and avoid this context rot, often referred to as “context engineering”. But that made me curious: is there anything we learn from these approaches, to improve our own personal context rot? Can we context-engineer our own personal effectiveness? Let’s explore!
1. The smallest possible set of high-signal tokens
For LLMs, this is really the guiding principle to avoid context rot and maximize the chance of receiving the desired outcome. Anthropic defines context engineering as “the art and science of curating what will go into the limited context window”, so achieving a high signal-to-noise ratio is clearly critical. They recommend writing prompts that are extremely clear with simple, direct language, and with ideas presented at the right altitude for the agent.
What inspiration can we take from this technique for our own personal context efficiency? Here are a few approaches that I can think of.
Keep your goals focused, then align your activities accordingly. Reduce the attention drain of activities that are not aligned with your goals wherever possible.
Single-thread your tasks. For me, multi-tasking is a recipe for personal context rot. It’s sometimes tempting to draft those user stories whilst half-contributing to that important team discussion, but chances are you’ll do a bad job at both. Focus on one task at a time, but choose your tasks wisely. Plan ahead to batch similar tasks together, to make the best use of your focus.
Remove distractions. According to a study, it takes an average of 23 minutes to regain your focus after an interruption. Another researcher has ascribed “attention residue” as the cause, on the observation that a portion of your brain remains attached to the first task when you switch to a second.
That reduction of your cognitive resources available for the second task sounds a lot like the human equivalent of context rot. So do whatever you need to eliminate those distractions and stay focused. My favourite distraction-buster strategy is to shut down Slack and email entirely for portions of my day, so I can stay focused on the task at hand.
2. Session resets and compaction
For LLMs, starting a new session is a guaranteed way to reset the context window, ready to handle a new set of information and instructions. But the downside is that resets, by definition, remove all the context that was built up in the previous session. So, AI agents typically employ some techniques to bridge the gap between sessions.
“Compaction” is one such technique, used to summarise the contents of a context window. It carefully chooses only the critical details to use in the next session. AI coding agents are increasingly built to record progress in incremental stages, for example via structured progress files and commits. An initialiser agent can then easily establish the context again in the refreshed session.
What can we humans do to achieve that attention reset, without losing track when we return?
First, take breaks. You need to give your brain a chance to process and reset. Relax, go for a walk, enjoy the company of others – anything to let that attention residue fade and let your brain recharge.
Externalise your memory through structured notes so you don’t have to hold facts and tasks in your memory.
Focus on capturing actionable insights, rather than the raw detail. What’s the actual opportunity that you might want to explore in the future? You’ve already done the thinking about why this piece of information is important, so don’t force yourself to have to remember that again in the future.
3. Just-in-time inference
Claude Code often refers to a lot of context data to achieve its tasks, but it doesn’t pre-process all that data up front. Instead, it dynamically loads reference data into context at runtime, from a set of stored references and indexes. That “just in time” approach means it can manage context more efficiently.
This feels like a natural fit with how our brains work, as we would certainly struggle to internalise and memorise large bodies of information. Here are a few personal context hacks inspired by just-in-time inference.
When is your best time for high-focus thinking? Optimise your day around that. My brain functions much better first thing in the morning, before I’ve taken on board a wide array of distractions. So whenever possible I keep my mornings clear for that focus work and schedule all my meetings in the afternoon.
Don’t pull at every thread. On a typical day, many more insights can crop up than you and the team can immediately act upon. Customer feedback, analytics insights, code refactoring opportunities, changes in the market, an idea from a stakeholder, a new piece of research, a niggling inconsistency in your product. If they aren’t aligned with your current goal, pulling that thread too far is a distraction. Note why it’s interesting and move on.
Rely on your system. Having a systematic way of taking notes is what frees you up to not pull at every thread. Tools like Notion or Obsidian are great for turning your notes into lightweight databases that let you tag and backlink items. But low-tech approaches can work great too – keep a structured to-do list handy, and organise those opportunities you uncover into an opportunity-solution tree, with sticky notes if you must!
Summing up
Clearly artificial intelligence and the brain of a product manager are not the same, but it was enlightening to think about the similarities between context rot, context switching and attention residue. I’m not sure we’ve uncovered any breakthroughs for fine-tuning our personal context windows, but perhaps that’s reassuring.
Given the rapid pace of progress in AI, Anthropic recommend to “do the simplest thing that works”. Whilst that advice relates to building AI agents, a similar approach to managing our own attention and brain energy rings true.
What’s your personal context-engineering hack? I’d love to hear how you protect your focus - drop a comment below.