Intro to Claude Context

Claude starts every conversation as a blank slate. No memory of your last chat. No idea who you are, what you do, or how you like to work. It only knows what you tell it.

Most people handle this one of two ways. They repeat themselves every chat, re-explaining their role and preferences from scratch. Or they cram everything into one endless conversation until the model starts losing track of what matters. Both feel productive. Neither is.

Long conversations are actively worse. They’re slower. And every instruction you add dilutes attention on the thing you actually need. More context is not better context.

To escape this, you have to understand how Claude’s context works.

Think in Tiers

Context lives in layers, sorted by scope, from always-on to one-shot. Four tiers. Each one has a job.

Tier 1: Global Instructions. Always on. Who you are, what you do, how you want to be talked to. Every conversation, every project. Keep it short. Three to five sentences. “I’m a project manager on the marketing team. I handle campaign planning and vendor coordination. Be direct and opinionated. Don’t give me several options when there is one obviously correct answer.”

That’s it. Claude reads this every time you send a message. You write it once.

Tier 2: Project Instructions. Rules for one workspace. Every conversation in that project gets them automatically. If you manage a monthly budget, the project instructions say the fiscal year starts in February, all figures are USD, format currency with commas, no decimals, and when you compare months always show the delta and percentage change.

The briefing you’d give someone joining the work. More specific than global. Less specific than a message.

Tier 2b: Skills. Recipes for how to do something. They load on demand when you need them. If there is one right way to do a task, or you keep explaining the same process, write a skill. A vendor email skill that always includes the account number in the subject, references specific PO numbers, and closes with a concrete next step. Said it three times? Write it down.

Tier 3: Your Message. This is where the work happens. Everything above feeds into it. “Compare October and November spend. Anything unusual?” Two sentences. Claude fills in the rest from context. It knows your tone preferences from global instructions. It knows the currency format, the variance threshold, the delta requirement from the project. You get a precise, consistent answer from a casual question because the structure was already in place.

Tier 4: The Fetch Layer. Stuff Claude goes and gets without being asked. It reads your context, realizes it needs something, and pulls from connected sources: databases, web search, internal tools, email. Fetched data carries more weight because of where it lands, at the bottom of context, right before Claude responds. The last word in any conversation matters most.

Memory Is Not the Answer

Claude has a Memory feature. It learns things about you over time so you don’t have to repeat them. But it’s unreliable. Claude decides what to remember, and its judgment doesn’t always match yours.

For anything that matters, write it down deliberately. Global instructions, project instructions, skills. Don’t hope the model remembers the right things at the right time.

The Prompting Decision

If you’d be annoyed when it changes, specify it. If you’d be delighted when it changes, say nothing.

This applies everywhere. Global instructions, projects, skills, messages. Currency format? Specify it. You don’t want that varying. The exact phrasing of an email greeting? Leave it alone. Let the model surprise you. Constraint where it matters, freedom where it doesn’t.

The trap is over-specifying. People write instructions for things they don’t actually care about, and every unnecessary constraint is attention the model spends on your rules instead of your problem. Don’t get in the model’s way. Precision matters, but precision is not the same as volume.

Getting Started

1. Write your global instructions. Three to five sentences. Who you are, how you work.
2. Create a project. Pick something you do weekly. Give it context.
3. Use Claude for a week. Notice when you repeat yourself.
4. Turn repetition into skills. Said it three times? Write it down.

Keep global instructions tight. Be specific where it matters, which is projects and skills. Let the message be simple. If your message needs a paragraph of setup, that setup belongs in a tier.

You’ll get better at this. Everyone does.

---

Converted from slide deck.