Context is a budget: five habits that keep AI output sharp
Claude uses the word "context" for two different things, and that can be confusing.
Project context is what you give Claude: the project files and background information covered in my last blog.
Session context is the working memory Claude actually has in one particular session.
When a session's working memory fills up, the quality of the output drops. The two are related. The more files Claude has to read to action a task, the faster the session fills. Everything counts against it: your messages, the files it reads, its own answers. Claude Code can compact a session automatically (autocompact), but I don't rely on it to decide which parts of my project matter. I decide.
If you use Claude Code, you can keep an eye on how much working memory you've used, and make sure key learnings from each session are committed to memory so they aren't lost for future work.
These are the five habits I use.
Check the meter. Type /context to see how much you've used. Look before a big task, not after things go wrong. There's also a way to have a visual circular red dial show up when you pass 50%, so you don't have to remember to look.
Summarise before you compact. Whenever the context gets above about 50%, I ask Claude to update Progress.md so any key learnings are committed to that file. The important parts then live in a file that survives, not only in a chat that gets compressed.
Compact at 50%. That's my rule of thumb. It isn't a law, so watch how your own sessions behave and adjust.
Start a new session with a handover prompt. Before I start the new session, I ask the existing one to write a structured handover prompt to continue the work in a new session. I paste that whole prompt into the new session and start again with a fresh window of context and fresh working memory. That way, the new session is fully up to speed with what has already been done, what has been learned, and what needs to be done next.
Send deep research to a subagent. A subagent has its own context (working memory). Using subagents to conduct literature searches, draft web pages or research particular topics is a good idea whenever you're primarily interested in the outcome of that research - because the heavy reading doesn't fill up the context of your main session. You just get the result.
Two commands are worth knowing alongside these:
/clear wipes the current session's context, so a new task starts clean.
/resume takes you back to an earlier session with its context, when you need it again.
Why does this matter for life science marketing work?
Because the tasks are long. A messaging framework, a set of web pages or a claims review can run across many steps. If the AI has quietly lost the rules you set at the start, you get an inconsistent result, or a lower-quality one, and the inconsistency is hard to spot.
The same goes for what it learned along the way. If that information drops out of the session's working memory, the quality of the output drops with it. And if the learnings were never written to a file, future sessions don't benefit from what this one worked out.
The habits above take a few minutes to learn. After that they cost almost nothing.
If you'd like help building this into your team's workflow, get in touch.
Next issue: skills and agents, and how to package a repeatable workflow.
Have you noticed AI output getting worse late in a long session?
Previously in this series: