Where you work, and the model behind it

A minute on each, and the rest is in the work itself.

The app is Claude Code, and you reach it on the web

The company works in Claude Code on the web — open it at claude.ai, the same place you already sign in.

The desktop app and the phone app are the same Claude Code; the web is the default, and your work follows you across all three.

Everything is already connected: you talk to Claude, Claude is inside your workspace, and the workspace holds your systems, your credentials, and your profile — set up for you, nothing to configure.

So you open it and start; the plumbing is done.

The bar above the chat — set once, then forget

Start a session and the bar above where you type shows the workspace you’re in, and carries over your last setup — so once it’s right, it stays right.

Two settings there are worth knowing:

• Permission mode — most of us run auto.
• I work straight through a task instead of stopping at every step to ask for a yes you’d always give; you stay in control by reading the result, not by gating each move.
• The outbound gate from the last module still holds — auto speeds the steps in between, it never sends on its own.
• Model — which Claude is doing the work, covered just below.

Seeing changes, reading, and sharing

On the web there’s no file panel open by default, so here’s how to see and move your work:

• See what changed. When I edit a document, the change shows inline — green for what’s added, red for what’s removed — so you can check exactly what moved.
• Read the whole thing. Say "paste the full document" and I drop the entire document into the chat to read.
• Share it. Say "publish this as a Google Doc" — I publish it through the workspace publisher, and from Drive you share it with anyone.

Four models — when to reach for each

You pick which Claude does the work from the model menu:

• Opus 5.5 — your default. Everyday work, building a new skill, deep research — Opus handles all of it, and it is the model Anthropic recommends starting on.
• It runs with the full 1M context on every plan, so a heavy session has room to run.
• Sonnet — for high-volume routine work. Half the price of Opus, for repetitive tasks where cost matters more than depth.
• Haiku — for speed. Fastest and lightest, for quick high-volume work like sorting a long list.
• Fable — for the hardest work. Not for everyday work; reach for it when Opus falls short, or when the work is new, ambiguous and expensive to get wrong.
• Effort sits beside the model and sets how hard I work; Appendix-7, Choosing the model and the effort, covers which setting fits which task.

When a session fills up — the starter prompt

Why it happens.

Every model holds a fixed amount at once, even with a 1M context, and a long session eventually runs low on room.

The move.

Don’t push through a session that’s starting to slip — carry the work to a fresh one with a starter prompt.

Say "write me a starter prompt to continue this in a new session," and I produce one that points the next session at three things:

• the tail of this conversation — the last stretch of the transcript, so it knows where we ended.
• the knowledge file we worked from, so it loads the same ground.
• the goal and the next task, so it knows what to do first.

Then carry it.

Say "save it in the chat as markdown I can copy," paste it into a fresh session, and we pick up exactly where we left off.

This is not obvious, and it’s the one move that lets long work survive a context limit — worth practicing once before you need it.

(That’s different from coming back to a session tomorrow, from the last module — this is for when a single session has filled up mid-work and you need a clean one to keep going.)

The one thing to remember

You work in Claude Code on the web, already connected to everything; run auto and check by reading; default to Opus 5.5, step up to Fable only when Opus falls short; and when a session fills up, ask for a starter prompt and carry the work onward.

Next: the craft levers — the power moves, writing that lands, and building your own expert.

Sources

• Claude Code on the web reference (verify against the live doc): https://code.claude.com/docs/en/claude-code-on-the-web — web at claude.ai, desktop and mobile apps, the workspace/settings bar, the permission modes, inline diffs.
• Claude model lineup (verify against the live doc): https://platform.claude.com/docs/en/about-claude/models/overview — Fable 5.1 (demanding reasoning and long autonomous work), Opus 5.5 (1M context, the recommended starting point), Sonnet (speed/intelligence balance), Haiku (fastest).