Vision: images, screenshots, charts and visual PDFs

0:00 Vision that you can trust

You can hand Claude a screenshot, a chart, a photo of a whiteboard or a stack of ad creatives, and ask what it sees. That is powerful, and it is also where some of the sneakiest AI errors hide. In this lecture you will learn what Claude's vision is good at, a three step pattern that prevents silent misreadings, and how to run a professional creative review.

0:29 Why vision needs care

Why does vision need special care? Because images feel objective. A chart is a chart, right? But when an AI looks at a picture, it is interpreting pixels, and a slightly blurry label or two similar colours can be read wrongly. The danger is that the error does not stay small. It travels silently into a conclusion that sounds perfectly reasonable. It is like a colleague copying a number wrong from a whiteboard. Everything they calculate afterwards is precise, and wrong. So the goal of this lecture is to make perception visible, so you can check it.

1:11 What vision covers

Claude can analyse photos, screenshots, charts, diagrams, handwriting, scanned forms and the visual pages of PDFs. Recent models got noticeably better with high resolution images, so dense dashboards and small text read more reliably. One distinction matters. Vision is analysis. Claude is not a photo generator. It can, though, create visual work as design and code, charts, diagrams, slide decks and visual mock ups, which you can then edit.

1:41 The silent misreading

Here is the problem. A blurry six gets read as an eight. Two similar legend colours get swapped. And then the model reasons perfectly, from a wrong starting point, and gives you a confident, wrong conclusion. The fix is to separate seeing from thinking.

2:00 Describe, then Answer, then Flag

So use this pattern every time. Step one, describe. List everything you see, every number, label, axis, legend and date, without interpreting. Step two, answer my question using only what you described. Step three, flag anything you were unsure about reading. If step one gets something wrong, you correct it before any conclusion is drawn. It takes seconds and it removes most vision mistakes.

2:28 Better inputs, better answers

Help Claude help you. Crop to what matters. Zoom in before screenshotting small text. Label multiple images, image one is March, image two is April. And if you actually have the data as a spreadsheet, upload the spreadsheet. Vision is for when you do not have the numbers in a cleaner form.

2:51 Simple example

Here is a simple example. Photograph a restaurant receipt and ask, how much VAT did I pay? Instead of letting Claude jump straight to a number, use the three steps. Describe, list every line, amount and total you can see. Answer, using only that list, what is the VAT? Flag, anything you could not read clearly. Claude lists the items, calculates the VAT, and flags that one faded line might be seven or one. You glance at the paper, see it is a one, and you are done. Thirty seconds, and you know the number is right.

3:33 Worked example: creative QA

A performance marketer at a UK e commerce brand uploads four ad variants and the brand guidelines. For each ad, Claude describes the layout, scores clarity at thumbnail size, brand fit, hierarchy and mobile legibility, quotes the guideline rule behind every deduction, and suggests one fix. It flags that one variant's price text fails the contrast rule, and another uses a retired logo. The marketer checks both against the files, fixes one and drops the other. And the overall ranking? Treated as a hypothesis. The A B test decides.

4:12 Privacy and ethics

Before you upload, blur or crop personal data. Customer names, emails, order numbers, browser tabs and notifications all leak easily in screenshots. Claude is designed not to identify real people from their faces, and you should not try to use it that way. And client creative belongs in your organisation's approved workspace, not a personal account.

4:36 Charts and dashboards

Dashboards deserve special care. First ask Claude to reconstruct the data from the chart as a table, then compare periods. And before anyone reports a trend, ask this question. Is this change large compared with the normal week to week variation you can see in the chart, and what would you need to be confident? That question has saved many marketers from announcing a breakthrough that was really just noise. Treat aesthetic scores on creative the same way, as hypotheses to test, not as results.

5:13 Turn it into a team asset

Once your describe, answer, flag prompt works well on your own creatives, turn it into a shared asset. Save it in your prompt library, and later in this course you will learn to package it as a Skill that your whole team can use, so every ad, landing page and social post gets reviewed against the same guideline rules. You will know it is working when misreadings get caught in the describe step rather than in a client meeting, and when feedback cites specific rules instead of personal taste.

5:52 Try this now

Try this now. Take a screenshot of a chart you actually use at work, from your analytics tool, an ad dashboard or a sales report. Crop it tightly around the chart and blur any customer names or account details. Send it to Claude with the three step prompt, describe everything first, then answer your question using only that description, then flag anything uncertain. Compare the described numbers with the real ones on screen. Did it misread anything? Write down one thing you caught or confirmed. Do this three times this week and you will develop an instinct for when vision needs a second look.

6:37 Watch me do it, part 1

Let me show you the creative review for the UK autumn sale. I drag in four ad variants and the brand guideline PDF. Before anything else, I label them. Image one is the price led version, image two the lifestyle version, image three the bold text version, image four the carousel cover. Then I paste the review prompt. For each image, describe the layout, headline, call to action and colours you see. Score one to five on offer clarity at thumbnail size, brand fit, hierarchy and mobile legibility. Quote the guideline rule behind any deduction. Suggest one fix. Finish with a ranking table and a list of uncertain readings. Then I send it.

7:26 Watch me do it, part 2

First, I read the descriptions before the scores. For image three it says the price reads twenty nine pounds, and I zoom in to confirm it did not misread a digit. Good. Then the flags. It says image three's price text fails the minimum contrast rule and quotes section four point two. I open the PDF and the rule is there. It says image two uses the old logo lock up, and page six confirms it. So I drop image two, ask the designer for a higher contrast price on image three, and re upload the fixed version for a quick recheck. The ranking table goes into my test plan as a hypothesis, and the A B test will decide.

8:18 Recap and next step

Recap. Vision reads almost anything visual, but separate perception from reasoning with describe, answer and flag. Feed it cropped, legible, labelled images, and upload real data when you have it. Blur personal information first. Your next step is to run the pattern on a real dashboard or ad from your work, note one misreading you caught, and save the best version as your team's creative review prompt.