Claude is one model behind a handful of interfaces, each with its own rules. The chat at claude.ai. The app on your phone. A terminal, the plain text window programmers live in, where it runs commands. A Discord bot that wakes up on its own.
What changes between them isn't intelligence. It's what Claude can touch: your files, the open internet, your apps, a browser, the clock. Pick the door that matches what the job needs to reach. Most people only ever use one.
Same Claude behind every door; the only thing that changes is what it can reach from where it stands.
Four doors here, and odds are you've only ever opened the first.
A plain chat packs nothing; a Project chat walks in with its bags already full.
This is the question that brings most people here: why did it forget?
Whatever you say lands in the smallest ring unless you put it somewhere that holds.
Stable facts: your role, your preferences, the tools you use.
Anything tied to one conversation. Scratch work, yesterday's half-finished draft.
Detailed factual recall. Memory is a gist, not a library.
Standing instructions and files for one topic, loaded into every chat you start there.
Projects don't share with each other. They don't sync with memory, and don't travel well to the phone.
When a topic outgrows its Project, or needs notes that live somewhere else.
Once the apps are plugged in, you stop carrying things to Claude and start asking it to go look.
Packaged instructions that run the same way every time. Reach for one when you keep typing the same preamble.
Turn on Slack, Gmail, Drive, Calendar in settings. The big unlock: stop copy-pasting threads into chat. Ask Claude directly.
Working pages, diagrams, documents, and small apps render next to the conversation, so you can revise without leaving the thread.
Everything above is the everyday layer. For most people it's plenty. Stop here and you'll still get real value tomorrow.
Want more control? The rest is the same model with more rope: the terminal, several Claudes at once, and work that runs while you sleep.
The matrix is the whole section; the fine print lives in the companion doc.
Chapter one covered chats and Projects. Here are the same doors from the engine room. A rented server is Claude on a box you pay for, running on a schedule. You pay for the box and the words.
| surface | your files | the open internet | your apps | a browser | the clock |
|---|---|---|---|---|---|
| chat & apps | |||||
| desktop app | † | † | |||
| phone | † | ||||
| claude code | † | † | † | ||
| a rented server | † |
Five doors, one question every time: what does this job need Claude to touch?
full capability fine print -> claude-power-setup.md
Every job sits somewhere between holding the wheel and handing it over.
Every rung you climb trades a little steering for a little more done while you're gone.
Pick the lowest rung that fits. The higher you climb, the more you have to spell out up front, and the less you can steer once it's moving. Fuzzy spec, stay low.
rung-by-rung commands -> claude-power-setup.md
Newer Claude models decide for themselves how long to think. There's no fixed budget to set. What you control is the nudge. Effort has settings, and the default is right for most work. The top setting mostly buys overthinking.
More effort helps until it doesn't; the last two notches are mostly spent second-guessing.
"Think carefully and step-by-step before responding; this problem is harder than it looks."
"Prioritize responding quickly rather than thinking deeply. When in doubt, respond directly."
the level names as the product shows them -> claude-power-setup.md
Follow it with a finger. The first yes wins.
Follow the questions down; the first yes is your door, and no four times lands you in a plain chat.
| Model | Feel | When |
|---|---|---|
| Opus | depth ●●● · speed ●○○ | The big jobs: coding, long agent runs, anything where being right matters more than being fast. The default for real work. |
| Sonnet | depth ●●○ · speed ●●○ | High-volume, cost-aware work. Loops, batches, light coding. |
| Haiku | depth ●○○ · speed ●●● | The cheapest path. Sorting, routing, quick classification. |
exact model ids and when to hand off -> claude-power-setup.md
| What goes sideways | The rule |
|---|---|
| Asking the phone to do big structural work | Hand it to a desk. The phone can't see your files. |
| Splitting work across helpers that didn't need splitting | Only fan out when the pieces are truly independent. |
| Editing the same files from two surfaces at once | Pick one. Parallel edits collide. |
| Maxing the effort setting for normal tasks | The default is the default for a reason. Max mostly buys overthinking. |
| Trying to set a fixed thinking budget | Fixed budgets are gone. Use the nudge phrases above instead. |
| Asking a sealed-off surface to reach the internet | Know what each door can touch. Package the job for a door that can, or run the command yourself. |
| Calling it done because the light is on | "Running" is not "working." Call the endpoint, check the output. Status is not correctness. |
It's a context system that travels with you. Drop raw sources in, get a structured wiki out, and stop re-explaining yourself to every new Claude session. Karpathy+ walks through how to build one, based on Karpathy's LLM Knowledge Base pattern and extended with ingestion loops, cross-project references, and quality guardrails.
Read the Karpathy+ guideI set systems like this up for individuals and small teams. If that's you, the discovery kit is where it starts.
A short email to a list I run myself. No spam, no selling.