Working with AI
What these tools actually are, why they slow down, and how to get noticeably more out of them.
Applies to Copilot and Claude alike. The tools differ; the way they work does not.
The words we'll use
None of this is complicated. The vocabulary is what trips people up.
- LLMThe engine under Claude and Copilot.
- PromptEverything you send, not just what you typed.
- TokenThe unit of measure. About four characters.
- ChatOne conversation, re-read in full each turn.
- ModelWhich version you pick. Bigger costs more.
- Context fileYour decisions, saved for the next chat.
- ProjectRelated chats, plus background they share.
- .md fileMarkdown — plain text with formatting.
- ConnectorA direct line into a system you use.
- MCPThe standard plug. Think USB.
What is an LLM?
The simplest way to picture it: your phone's autocomplete, scaled up enormously. It doesn't look answers up — it predicts them.
It keeps patterns, not pages — so it always produces something, even when it's wrong.
What is a prompt?
It's everything you send to the LLM — not just the line you typed.
You write one line. All of it is the prompt.
Every message re-reads the whole conversation
There is no memory between messages. It starts from the beginning, every time.
What is a token?
It's the unit of measure — everything read and written is counted in tokens.
Both are counted. A long chat spends most of its budget re-reading.
A file you attach is read again every turn
You attach it once. You pay for it every turn — and a pasted screenshot counts too.
Keep each chat to one subject
Everything in the chat gets re-read — including the parts that have nothing to do with your question.
Bigger models think harder and cost more
Same kind of engine, different sizes. The cost is per token — in and out.
Prices are per million tokens — so the model you pick moves the bill five-fold.
Match the model to the job
| When you're doing this | Reach for | Because |
|---|---|---|
| Reasoning across several documents | Opus 5 | Real judgment, and the answer has to hold up |
| Everyday drafting and questions | Sonnet 5 | What most work actually needs |
| Quick rewrites and short emails | Haiku 4.5 | Quality isn't the constraint here |
| Follow-ups, once the thinking is done | Drop a level | Don't pay twice for the same reasoning |
Being economical on the one task where quality shows is a false saving.
Keep the thinking, drop the bulk
Nothing new — it's the role and project reference docs we've always written, so the next person reads the memo instead of the whole history.
How to get a summary: just ask
A project shares one file with every chat inside it
Every new chat starts already knowing the background. You stop re-uploading.
Markdown is just a few symbols
On the left, what you type. On the right, what those symbols turn into.
Kitchen Remodel
Decisions
- Shaker doors, white
- Quartz counters
Budget: $32,000
Plain text that still knows what's a heading and what's a list — nothing to learn beyond a handful of marks.
Save working notes as markdown
Exactly the same notes, saved four ways. The bar is the file size — what the tool has to read through before it reaches your words.
What this looks like in practice
Now swap kitchen remodel for a client. Same structure, exactly.
Three things to try this week
Summarise a chat into a file
Ask for the decisions, the numbers, and what it got wrong. Save the .md.
Group related chats into a project
Gather everything about one piece of work in a single place.
Add context every chat shares
Drop that file — or a spreadsheet — in. New chats start already knowing it.
Advanced
Where this is heading — connecting the tools to your systems, how Copilot and Claude actually differ, and working straight off your own files.
Connectors — one plug into everything
A direct line from the tool into a system you already use. MCP is the standard plug shape that makes it possible — think USB.
Copilot is the machine. Claude is the brain.
Copilot wraps a model — the brain — and connects it to your Microsoft 365 through a layer Microsoft calls Work IQ.
Microsoft now runs Anthropic's models inside parts of Copilot — so the brain in the slot may be Claude either way.
Claude chat vs Cowork
Chat works over what's in the cloud. Cowork also reaches the files and folders on your machine.