Lesson 07· Module 7 — Surfaces· 40 minutes, one sitting

Real Material In, Real Work Out

Four ways material moves in and out of a conversation, and three controls underneath it that almost nobody touches. This is the tightest sitting in the course — the most to do per minute — so the live half is timed.

Six lessons in, everything you've learned has been about what happens inside a conversation. This one is about its edges: what you can push in, what you can pull out, and the settings that decide how much work gets done in between. It's the least conceptual lesson here and the most immediately useful — and it closes a loop opened on the first page.

In · Give it the actual file

You can attach documents, spreadsheets, PDFs and images directly to a conversation. That single move does the thing Lesson 5 spent thirty minutes on: it takes a claim from unsupported model knowledge to a source you supplied — the top of column A. You can go and look at what the answer rests on.

Which is a real gain, and it is also where the most dangerous assumption in this lesson lives.

The thing to unlearn

“I uploaded it, so it's read the whole thing.”

A successful upload tells you the file transferred. People read it as a guarantee of complete, accurate reading, and then trust a claim drawn from deep in a long document exactly as much as one from the first paragraph. This is worse than an ordinary mistake, because it manufactures confidence at precisely the moment Lesson 5 taught you that traceable material is better material.

The principle

“Uploaded” is not a single file-reading mode. How completely a file gets read depends on its format and its size, and you cannot tell which you got by looking at the reply.

A PDF, a spreadsheet and a photograph of a page are not handled the same way. A short file and a long one are not handled the same way. The reply reads identically either way — same fluency, same specificity, same absence of any hedge. Which is the same signal-free surface you met in Lesson 1, arriving through a different door.

An earlier revision of this lesson taught page thresholds, and they've been cut. The numbers came from platform documentation describing a different path than the one you're using, and were never verified against it — exactly the sort of claim Lesson 5 says not to accept. The deeper problem is that a named number invites you to treat it as an operating parameter, so when it changes it fails twice: the number is wrong, and the habit built on it was wrong. What survives every future change is format and size both matter, and you can't tell by looking.

You don't need the mechanism. You need the conclusion, and it's one you already own:

The rule this lesson installs
If a claim from a file matters, demand the supporting line — and then go and find that line in the file yourself.

Two seconds of asking, ten seconds of looking. Note that this is a check, not a trap: a quoted line that matches the source is a good outcome and tells you something. The habit is worth having because it costs almost nothing and works the same whether the file was read completely or not.

Out · Artifacts

When what Claude produces is substantial and self-contained — a document, a table, a piece of code, something you'd plausibly want outside this chat — it appears in a separate panel beside the conversation rather than inline. That's an Artifact.

Most casual users have seen one and never used either of the two things that make it worth caring about.

1 · It can be changed in place

You don't have to ask for the whole thing again. A targeted request — “cut the recommendations section to three lines”, “add a row for units shipped” — asks for a change to the document that already exists rather than a new one, and earlier versions stay reachable in the panel. Watch which parts move and which stay put.

This matters more than it sounds, because the alternative is the reroll trap wearing a suit. Regenerating a whole document to fix one paragraph re-draws every other paragraph too, and you can lose three things you liked to fix one you didn't. Targeted change request in, targeted change out.

2 · Publish makes a link

Publish turns an Artifact into a web page that anyone holding the link can open, with no Claude account and no sign-in. Say that to yourself plainly, because the word does a lot of quiet work: publish here means publicly reachable by whoever has the URL. Not shared with people you named. Not visible to your team only. Anyone with the link.

That's exactly what you want for a one-pager you're sending to a client, and exactly what you don't want for anything containing a real customer name, an unreleased number, or someone's salary. The control is fine. The default assumption people carry into it is not.

On Free, Pro and Max the button is Publish and the link is public. Team and Enterprise accounts get a Share route that keeps an Artifact inside the organisation instead. If you're on a work account and can't find Publish, that's why — and it's worth knowing which one you have before you send anything anywhere.

Alongside · Branching

Editing one of your earlier messages doesn't overwrite the conversation. It creates another version of it from that point on, with its own replies and its own Artifacts, and you can move between versions.

Worth being precise about what this is: it's a property of the conversation, not a feature of the Artifact. It happens to be useful while you're developing one — try a much shorter brief without losing the long version, keep both, compare — but the mechanism is the message you edited, not the document beside it.

It's also the one thing in this lesson you can skip today and lose nothing. Nothing in the capstone depends on it, and forty minutes is not much. It sits at the end of the exercise as an optional extension for that reason.

Both · Connectors

Google Drive, Gmail and Calendar can be connected to Claude, after which it can bring that app into a conversation when doing so fits what you're asking — pulling a document out of Drive, reading a thread, checking what's on Thursday.

Two things follow that are worth holding onto. A document pulled from Drive stays synced to the live version, so it's a different kind of source from a file you uploaded once and forgot about; the upload is a snapshot, the connector is a pointer. And on Pro you connect these yourself — there's no administrator step in your way.

Deliberately not promised: that a connector fires only when a request strictly requires it. An earlier draft of this lesson said that, and it over-claims — it describes routing as more deterministic than anyone documents it to be. In a course about knowing what your output rests on, promising false precision about when a tool activates is a poor lesson. Assume "when it fits what you're asking", check what a given answer actually drew on, and connect only what you're comfortable having available.

Connectors are a parallel path here, never a gate

A meaningful share of people won't connect a personal Gmail for a course exercise, and that's a reasonable position rather than a gap in commitment. The exercise below runs identically over uploaded files. Both routes satisfy the objective and both count. If you'd rather not connect anything, upload the files and carry on — you're not doing a reduced version.

The dials underneath

Click the model name at the top of the conversation. What opens is not a model picker — it's three separate controls, and casual users almost never touch any of them.

ControlWhat it's forTurn it up when
Model Which model answers. The default is built to handle most of what you'll throw at it. The analysis is genuinely hard, not merely long.
Effort How much work goes into a response. More effort costs time and usage; less stretches both further. The task is hard, or being wrong is expensive.
Thinking A separate toggle giving room to work through a problem before answering. Combines freely with any effort level. Anything multi-step, or anything worth getting right.

On model: learn the dial, not the lineup. Model names change often enough that memorising the current set is a way of learning something with an expiry date on it. The durable version is default for most things, step up for genuinely hard analysis, and you can apply that to a menu you've never seen before.

On effort: the same reasoning, which is why this lesson names no tiers. They have names, the names will change, and what won't change is the trade — more effort buys thoroughness and costs you time. You'll also see options aimed at long-running coding and agentic work, which are out of scope for this course. Knowing which part of a menu you can ignore, and why, is itself the skill.

On thinking: remember Lesson 1, and the fiddly multi-step calculation that came out better when it was asked to work through the problem one step at a time? That's this, as a setting, applied without you having to phrase it. It's the same lever with a switch on it.

Reach for the prompt before the dial

Lesson 4 named seven repairs and deliberately left the seventh — some failures are a settings problem, not a prompt problem — until now. Here's why it waited. A settings change is the easiest move available: one click, no thinking, feels like progress. So it's the move people reach for first, and most of the time it's aimed at the wrong thing.

More effort on an under-specified request buys you a more thorough version of the wrong answer. A stronger model still can't know which parts you cared about. Check the specification first — task, context, input, output shape, constraints, examples — and reach for a dial when the brief is already complete and the work itself is genuinely hard. Then the dial has something to work on.

Try it yourself

One run, in five steps and an optional sixth, against supplied material — a Lumen & Co. returns spreadsheet and the memo it arrived with. Minute budgets are on each step because this sitting is tight; treat them as a pace, not a rule. As always, record what actually happened, including “it handled all of that fine” — that outcome is information, not a broken exercise.

Open the material →  Read the memo there, then download both files from that page.

Do this — Tier B, about 22 minutes live

1 · Get it in. 4 min. Start a fresh chat. Attach lumen-q3-returns.csv and lumen-q3-memo.txt, and ask for nothing yet. Or, the parallel route: put both files in Google Drive, connect Drive, and ask Claude to work from them there. Either route is the exercise.

2 · The check, before you trust anything. 5 min. Ask a question whose answer sits deep in the table — for instance, which product had the worse return rate in the week ending 15 August, and what were the two figures behind it. Then demand the exact row it used, quoted. Then open the material page and find that row with your own eyes. Record whether it matched.

Then, still before trusting any total: that sheet contains at least six separate things that would corrupt a naive sum. Ask what they are. Name the ones you find yourself. Compare the two lists — the disagreements are the interesting part in either direction.

3 · Make the deliverable. 4 min. Ask for Dana's one-pager — what's driving returns, which line is the real problem, one defensible recommendation — as an Artifact. Note where it lands, and what about that content fits the trigger: substantial, self-contained, and something you'd want outside the chat.

4 · Change it twice, in place. 4 min. Two targeted requests, not two regenerations. Something like “cut the recommendation to three lines”, then “add the units-shipped figure to each row of the table”. Watch what changes and what doesn't, and find the version history.

5 · The dial comparison, then publish. 5 min. Take the hardest analytical question this data supports — a good candidate is which product line genuinely deteriorated across the quarter, and which apparent deterioration is an artefact of how the numbers were recorded. Ask it on defaults. Ask it again in a fresh chat with thinking on. Compare, and be honest about whether the wait bought you anything: on an easy question it won't, and noticing that is the point of the step. Then publish the Artifact and open the link in a private window, so you see exactly what a recipient sees.

6 · Optional extension — branch. If you have time, and skip it freely if you don't. Go back to the message where you asked for the one-pager, edit it to ask for something markedly different, and send. Check whether you now have two versions of the conversation, each with its own Artifact, then move back to the original and confirm it survived.

Checklist — two binary things, self-verified

Not graded, and both are yes/no. One: the published link opened in a private window with no sign-in. Two: the Artifact was changed twice in place, and the earlier versions are still reachable. If either is a no, that step didn't happen — go back rather than ticking it.

Six boxes — five steps plus the optional branch. Tap as you go; it saves in this browser only.

Check yourself

You upload a long supplier contract and ask about the notice period. The answer names a specific clause and a specific figure, and the figure matters. What now?
Format and size both affect how completely a file is read, and nothing in the reply tells you which you got. The repair doesn't depend on knowing: demand the line, then go and look at it.
You asked for a client update and got something competent, well-written, and not what you wanted — wrong emphasis, twice the length you needed. Your next move?
Competent-but-wrong is the signature of an under-specified request, and it's the case where a settings change feels most like progress while changing nothing that matters.
You asked for a three-part analysis of a spreadsheet. The third part is missing entirely and the first two are thin. You're deciding what to change.
Both dial items in this quiz resolve to the prompt, and that ordering is deliberate rather than a claim that dials don't work. They do — when the specification is already complete and the work is genuinely long or hard. Check the brief first, then turn the dial.

From memory, without scrolling up: name the four surfaces this lesson covered and the three dials, then write out the sentence about uploaded files and the one thing to do about it.

Surfaces: files in; Artifacts out, which can be changed in place and published to a link anyone holding it can open without an account; branching alongside, which is a conversation capability rather than an Artifact feature; and connectors both ways, with Drive documents staying synced.

Dials: model, effort, thinking — and the prompt gets checked before any of them. The sentence: "uploaded" is not a single file-reading mode; how completely a file gets read depends on its format and its size, and you cannot tell which you got by looking at the reply. So: if it matters, demand the supporting line and go and find it.

Surfaces & Dials Lesson 07

"Uploaded" is not one file-reading mode.

Format and size both matter, and the reply won't tell you.
Publish = anyone with the link, no account needed.

Demand the line. Fix the prompt before you touch a dial.

Primary source for this lesson

Anthropic, Change the model, effort, and thinking settings — Claude Help Center. support.claude.com → change the model, effort, and thinking settings

Read it for the part this lesson deliberately withholds: the current effort tiers, by name, with what each is for. That page gets updated when they change and this one doesn't, which is the whole reason the names live there rather than here — and treating a help page as the current-state source while the lesson holds the durable part is a habit worth generalising. Skip the sections on the long-running coding and agentic options; you'll know them when you see them, and they're not what you're here for.