Context & Compaction

Co-Work inherits Claude's 200,000-token context window per session. That sounds like a lot — and it is, for a single-threaded conversation. But when you are running four parallel agents doing web research, producing analysis documents, and synthesizing findings into a 12-slide PowerPoint, you consume context faster than you think.

Once you hit roughly 90% capacity, Claude auto-compacts. It summarizes the conversation history to make room for new content. This is a smart, necessary mechanism — without it, the session would simply stop accepting input. But compaction is lossy. It keeps the high-level summary and drops the details. And here is the problem: you do not know what it kept versus what it dropped.

Maybe it kept the final recommendation but dropped the specific data points that supported it. Maybe it kept the slide titles but forgot the nuanced talking points you discussed for slide 7. Maybe it summarized four agents' research into a single paragraph, losing the granularity that made the research valuable in the first place.

The practical fix is straightforward, and it is the same fix that experienced developers use in Claude Code: save progress to files early and often.

Before your context gets bloated — not after, before — ask Claude to save intermediate results. "Create a markdown file with everything you found about competitor pricing — the full details, not just the summary." "Save the SEO audit results as a standalone document in the working folder." "Write the research findings to a file before we start building the presentation."

Files persist beyond compaction. If the conversation loses context — if Claude compacts and drops the details of Agent 2's research — the file is still there. You can reference it by name: "Read the competitor-pricing.md file and use those data points for the pricing slide." The file is your safety net. The conversation is ephemeral. The files are permanent.

This is also why prompt queuing is so valuable. When you queue your next prompt early — while context is still fresh — it gets processed before compaction has a chance to reduce your working memory. The lead agent has all four research outputs in context when it builds the synthesis, rather than working from a compacted summary.

Let me give you the tactical playbook for managing context in a production Co-Work session. This is the part most people learn the hard way.

First: save deliverables to files as soon as they are produced. Not at the end. As they come in. Each agent should produce a file, not just a conversation response.

Second: if you are planning a multi-step workflow — research, then synthesis, then deliverable — queue all the prompts early while your context budget is largest. The first step has the most context available. Use that advantage.

Third: after compaction, do not assume Claude remembers specifics. Test it. "What were the three key findings from Agent 2's research?" If the answer is vague, point Claude to the saved file. This is not a Claude limitation — it is a context window limitation that applies to every LLM on the planet.

Fourth: for long-running projects, consider breaking work into multiple sessions. Session 1: research and save findings to files. Session 2: load the files and build the deliverable. Each session starts with a fresh 200K window. That is twice the effective context for the price of one deliberate save-and-reload step.

Context management is not glamorous. It is plumbing. But the difference between someone who manages context proactively and someone who hits compaction by surprise is the difference between a polished deliverable and a deliverable that mysteriously lost half its detail in the last mile.