Guide
Context engineering for your workday
Context engineering is the practice of giving a model what it needs before it answers. It's usually discussed for building AI products, but the same idea applies to how you use ChatGPT and Claude every day.
What it means
The term caught on in 2025 among people building AI agents. Their observation was that most bad answers come from missing information, not from badly worded prompts. So rather than polishing the question, you work on what the model can see when it answers.
If you use ChatGPT or Claude at work, you already do this by hand. Pasting an email thread into a chat counts. The trick is doing less of it.
The context in a workday
Think about what a capable colleague would need to help you right now:
- What you're working on, and why
- What was decided recently, and by whom
- What was said on the relevant calls
- What you've already tried
- What's still open
Almost none of that is in a chat window. It's spread across calls, documents, Slack, your browser, your terminal and other AI chats. That's the gap.
Four layers to set up
1. Standing instructions
Who you are and how you like answers. Custom instructions in ChatGPT, preferences and memory in Claude, CLAUDE.md or AGENTS.md for coding agents. Write these once and revisit them monthly.
2. Reference material
The documents a piece of work depends on: specs, style guides, pricing. Put them in a ChatGPT or Claude Project.
3. Live context
What happened recently. This changes every day, which is exactly why it's usually missing. It needs to be collected automatically, or it won't be collected at all. Cachejay does this on a Mac: it records screen text, calls and agent sessions and organizes them by project, and agents read it over MCP.
4. Actions
Once the model has context, let it act through connectors: post the update, file the ticket, log the call. More on this.
Habits that help
- Ask for what you want, not for what to paste. “Draft the update for Acme from this week's calls” works when the model can look up the calls.
- Keep one place per project. One Project or one folder, so context doesn't fragment.
- Prefer current over complete. Three up-to-date files beat thirty stale ones.
- Check what the model used. Ask where an answer came from. If the context was wrong, fix it at the source so the next answer is right too.