You’re mid-session with Claude or ChatGPT. The work is good – you’ve built up context, made decisions, established constraints. But the conversation is getting long and the AI is starting to drift.
You know from experience: push through and the quality keeps dropping. Start a new chat and you lose everything you’ve built.
A context capsule solves this. It’s a compressed summary of what matters – packaged so a new session can pick up where this one left off.
I use these constantly. They take two or three minutes and save twenty minutes of re-explaining.
How to create one
Two approaches. Both work.
Ask the AI to write it. Before the conversation degrades too far, say something like: “Create a context capsule – summarise what we’ve discussed, decisions made, current state of the work, and anything that should carry forward.”
The AI knows what you mean. It’ll produce a structured summary. Review it – sometimes it includes too much or misses something important. Edit, then copy the output.
Write it yourself. If you know exactly what matters, a manual capsule can be tighter. Five things to capture:
- What we’re working on (the task or project)
- What’s been decided (so the new session doesn’t revisit settled ground)
- Where we are (current state – what’s done, what’s next)
- Constraints or preferences (things the AI needs to respect)
- What’s unresolved (open questions to carry forward)
That’s it. A few paragraphs. Doesn’t need to be polished – it needs to be accurate.
How to use one
Open a new chat. Paste the capsule. The AI reads it and you’re working again – with the context that matters, without the accumulated conversation that was degrading performance.
If you’re working in a project, paste it as the opening message. If you’re in a regular chat, same thing – the capsule is your starting context.
When you need one
Not every conversation needs a capsule. Short, self-contained chats are fine without.
You need one when:
- The conversation has been running long and quality is visibly dropping
- You’ve seen a compaction message (“this conversation has been compacted”)
- You’re about to switch to a different phase of work – research to writing, planning to execution
- You want to continue work in a different tool or project
- You’re ending a session but need to pick up the same thread tomorrow
When you don’t
If you’re asking a quick question, getting a one-off draft, or doing something that doesn’t build on previous context – skip it. The overhead isn’t worth it for conversations that don’t compound.
What a capsule looks like
Here’s a real one I created during a content development session:
Context Capsule: Content Strategy Session
Working on: Blog post series for prompt engineering pillar
Decisions made: Series structure confirmed (5 posts). Post 1 published. Post 2 brief approved – covers system prompt anatomy using a simplified protocol as the example.
Current state: Post 2 ready to draft. Need to create the example protocol first – must be real enough to pass operational truth but simple enough to be instructive.
Constraints: Endless Customers lens active – all three behavioural constraints apply. First-person voice. Show version history or iteration evidence when discussing systematic approaches.
Open questions: Where does “anatomy of a protocol” sit in the series? Possibly a new Post 3, bumping others down.
That took about a minute. The next session started at full speed instead of spending fifteen minutes establishing what I’d already figured out.
Want to understand why conversations degrade in the first place? How AI Context Actually Works explains the mechanics.
