Start with what you already use — chats, Projects, memory. Ramp into surfaces, modes, and effort levels when you're ready to push further.
current as ofApr 2026modelOpus 4.7
How to use this sheet
The first half (01–04) covers what most people see — the chat, Projects, memory, connectors, and skills. The second half (05–12) is the power-user layer: where Claude actually runs, how hard it thinks, and how much autonomy you can hand it. Read straight through on your first pass. Come back to specific rows as references.
01What Claude Actually Is
Claude isn't one thing. It's the same model behind a handful of interfaces, each with its own rules. The chat at claude.ai. The app on your phone. A terminal where it runs shell 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, your browser, your apps, the network, the clock. Pick the interface that matches what the task needs to touch. Most people only ever use one.
01in a chat
02on your phone
03in a terminal
04inside your apps
05on a schedule
02Where You Already Use It
CHATS
claude.ai · app
One thread, one task. Fast, disposable, no setup. This is where most people start and most people stay.
USE FOR
Drafting, explaining, brainstorming, quick research, one-off analysis.
context resets each new chat
no files unless you upload
PROJECTS
claude.ai · scoped
A chat bucket with persistent instructions and a file stash. Every new chat inside a Project starts pre-loaded with that context.
USE FOR
Ongoing topics: one client, one codebase, one book, one thesis. Anywhere you'd re-paste the same context twice.
context doesn't leak between Projects
doesn't follow you to mobile cleanly
MOBILE
iOS · Android
Claude on your phone. Thin context, no filesystem, subset of connectors. Account-level memory travels with you.
USE FOR
Inbound triage, short drafts, voice-to-text, thinking out loud on the go.
don't ask it to read local files
hand structural work to desktop
CONNECTORS
Slack · Gmail · Drive · Linear
Turn Claude into a reader and writer for the tools you already use. Enable one in settings and Claude can search, fetch, and act inside it.
USE FOR
"Summarize this Slack thread", "draft a reply to that email", "find last week's Drive notes", "update this Linear ticket."
scoped per workspace / account
privacy: Claude sees what it reads
03Memory & Projects · What Sticks, What Doesn't
Auto Memory
opt-in · account-scoped
What it keeps
Stable facts you've told it over time: your role, your preferences, recurring projects, tools you use. Things worth remembering next session.
What it drops
Anything session-specific. Temporary state. The current conversation's scratch work. Yesterday's half-finished draft.
Where it's thin
Cross-topic references. Long-term citations. Detailed factual recall — memory is a gist, not a library.
Projects
explicit · topic-scoped
What you get
Shared instructions and a file stash scoped to one topic. Every chat in the Project starts with them pre-loaded — consistent framing, consistent references.
What you don't
Transfer between Projects. Sync with auto memory. Travel cleanly to mobile. Access from Code or bots.
Where they break
When a topic grows past a single Project, or spans surfaces, or needs to reference work living in a different bucket.
When both cap out
You're looking for a knowledge base. Something Claude can read from any surface — mobile, desktop, terminal, bot — with the same context. That's what Karpathy+ covers at the bottom of this page.
04Your Toolkit · Skills, Connectors, Artifacts
Skills
Named behaviours you invoke on demand.
Packaged prompts with a predictable shape: /research, /brainstorm, /review, /summarize. Same input structure, same output structure, every time.
Use them when you find yourself typing the same preamble twice. Many ship with Claude; you can also install community and personal ones.
Connectors
Read and write inside apps you already live in.
Enable Slack, Gmail, Drive, Linear, Calendar, and more in settings. Claude can search, fetch, draft, and (where you allow) take action.
Biggest unlock for most people: stop copy-pasting threads into chat. Ask Claude directly.
Artifacts
Runnable output in a side panel.
Working HTML pages, SVG diagrams, React components, documents, interactive code — all rendered live next to the chat. Iterate without leaving the thread.
If you've seen Claude render a mini-app alongside your conversation, that's an Artifact.
chapter break
ADVANCED USE
Everything above is the visible layer. Below is where Claude actually runs — the surfaces, modes, and levels that let you drop into the terminal, spawn subagents, or let an autonomous loop run overnight. Same model, more control.
05Where Claude Runs
CODE
CLI · terminal
Claude Code. Full shell access, unrestricted network, file-system aware. Hooks fire here. Runs on your laptop.
USE FOR
System ops, deploys, multi-file refactors, SSH + git workflows, agentic loops, subagent fan-out.
no MCP connectors by default
no browser automation
DESKTOP
claude.ai · app
Claude in a browser or desktop app, sandboxed with code interpreter, MCP connectors, and file handling. Projects, skills, artifacts.
USE FOR
Research, analysis, content work, design, MCP-heavy tasks, connector-driven workflows.
sandboxed: no arbitrary network
no direct host-machine access
MOBILE
iOS · Android
Claude app on phone. Thin context, no filesystem. Subset of connectors. Cross-session memory via account.
USE FOR
Inbound triage, short drafts, connector queries on the go, thinking partner for quick decisions.
don't reason about local files
hand off structural work to desk
BRIDGES
3rd-party integrations
Claude behind a chat app: Discord bots, Slack apps, iMessage relays, custom webhooks. Built via the API.
USE FOR
Quick Q&A, reminders, push-notification workflows, team-channel assistants, always-reachable endpoints.
no long-session context
no wiki / knowledge base by default
SELF-HOST
your server · always-on
Your own VPS or server running the API on a schedule. Cron jobs, heartbeat scripts, scheduled briefs, webhooks. You pay for the box and the tokens.
USE FOR
Cron-driven agents, production bots, heartbeat services, scheduled digests, long-running integrations.
06Mode Ladder · How You Drive Claude
LOW ← AUTONOMY → HIGH
01
Interactive
Normal chat. Back and forth. You course-correct at every turn.
all surfaces
02
Skill / Slash cmd
Packaged behaviour invoked by name: /review, /plan, /analyze. Predictable shape.
all surfaces
03
Subagent fan-out
Spawn N workers in one turn for independent items. Results returned to the main thread.
Code · Desktop
04
Auto mode
Shift+Tab — long single task with full spec up front. Claude drives.
Code · Max tier
05
Timed loop
/loop 5m /foo — fixed cadence. Polls or checks on a schedule.
Code
06
Autonomous loop
/loop with no interval. Claude self-paces via ScheduleWakeup.
Code · Desktop
07
Scheduled
Cron triggers, remote agents fire on their own. No human in the loop.
Self-host · scheduler MCP
Pick the lowest rung that fits. Higher rung = more context to front-load, less ability to course-correct mid-task. Fuzzy spec? Drop to rung 1 or 2.
07Opus 4.7 Effort Levels
low
fast · cheap
medium
cost-aware
high
concurrent
xhigh
most work
max
hard only
Toggle mid-task to manage spend. max tends toward diminishing returns and overthinking.
Opus 4.7 uses adaptive thinking — no preset budget; Claude allocates reasoning where it's actually useful.
↑ want more thinking"Think carefully and step-by-step before responding; this problem is harder than it looks."
↓ want less thinking"Prioritize responding quickly rather than thinking deeply. When in doubt, respond directly."
Drift between global, account, and project-local context is the #1 silent failure mode. Keep one source of truth and treat the rest as views.
11Anti-Patterns · What To Avoid
Anti-Pattern
Rule
Asking Mobile to do structural or multi-file work
Hand off to Desktop or Code. Mobile can't see your files.
Spawning a subagent for work the model can already see
Only fan out for multi-file reads or genuinely independent items.
Running parallel structural changes in Code + Desktop
Pick one. Concurrent edits on the same files will collide.
Using max effort for normal tasks
xhigh is the default. max has diminishing returns and overthinks.
Relying on a fixed thinking budget
Not supported in 4.7. Use adaptive-thinking nudge phrases instead.
Attempting outbound network calls from a sandboxed surface
Know each surface's network limits. Package the request as a Code handoff or give the human the command to run.
Claiming "verified" from a process-status check alone
Call the endpoint, hit the function, check the output. Status ≠ correctness.
12Custom Status Line · Session Awareness
Claude Code supports a custom status line that fires on every session. Encode what Claude should know at-a-glance: which surface, which project, which effort level, whether your context file loaded. Keeps the model from having to re-derive state each turn.
current working directory (or NA if the surface has no FS)
project
personal · client-1 · client-2 · work · unknown
context
loaded if your CLAUDE.md / knowledge base initialized, missing otherwise
model
which model is handling the session (useful when swapping between Opus / Sonnet / Haiku)
A status line is not required — but it's the simplest forcing function against silent context drift. See Claude Code docs for how to configure settings.json.
Next level · context systems
The biggest unlock isn't prompt tricks.
It's a context system that travels with you. Drop raw sources in, get a structured wiki out, stop re-explaining yourself to every new Claude session. Karpathy+ is a full walk-through of how to build one, based on Karpathy's LLM Knowledge Base pattern — extended with ingestion loops, cross-project references, and quality guardrails.