IntermediateLesson 17 of 306 min read

Images and Screenshots as Input

Learn what images are good for, how to prepare a photo or screenshot so Claude can read it, and how shrinking and cropping change text size, with real measurements.

In this lesson you will

  • Know what images are useful for
  • Prepare photos and screenshots
  • Crop before you shrink
  • Spot what Claude may misread

About 22 minutes

A photo of a whiteboard that Claude could not read

Tom takes a photo of the whiteboard after the volunteers’ meeting, squeezes it down to send quickly, and uploads it to Claude: ‘Turn this into a task list.’ The reply is full of gaps and guesses. Half the writing was too small to read. Claude did not do a bad job. It was handed a bad picture. Images are one of the most useful things you can show Claude, and one of the easiest to get wrong. This lesson covers what images can do, how to prepare them, and what to watch out for.

What images are good for

  • Reading text in a photo or screenshot: a notice, a form, a menu, a whiteboard.
  • Explaining a chart or diagram: what it shows and what to notice.
  • Describing what is in a picture: objects, layout, mood.
  • Reviewing a design or a page: what is unclear or cluttered.
  • Helping you solve a problem on screen: a screenshot of an error message or a settings page.

Anthropic’s upload help lists the image types Claude supports as JPEG, PNG, GIF and WebP, and you can attach them, drag them in or paste them from the clipboard. Its tip for best results is to use images that are reasonably large, around a thousand pixels on each side or more, and to avoid small, low-resolution pictures. There are also limits on size and dimensions, which differ by plan and change, so check the current page before you upload something huge.

Prepare the picture

  1. Take it straight on. Hold the camera square to the page, in good light, with no glare or shadow.
  2. Fill the frame. Crop away the table, the walls and the thumb.
  3. Keep the text big enough. If you can barely read it yourself, so will Claude.
  4. Do not over-compress. Squeezing a file to make it small can destroy the thin strokes of small text.
  5. Cover private details. Black out names, addresses and numbers you do not need to share.

Numbers behind the advice

The program below makes a large screenshot of a rota, 3000 by 1800 pixels, with body text 40 pixels tall. Then it makes three smaller versions, resizing so the longest side is 3000, 1600 and 800 pixels, and reports the new size, the file size in kilobytes and how big the body text becomes after shrinking. It marks text under 14 pixels as small, which is our rule of thumb for text that is hard to read. Finally, it crops the image to the top left quarter, to show how much you can save by keeping only the part that matters. The picture is made by the program, so every number is real.

import os
from PIL import Image, ImageDraw, ImageFont

def font(size):
    for p in ("/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf",):
        if os.path.exists(p):
            return ImageFont.truetype(p, size)

shot = Image.new("RGB", (3000, 1800), "white")
d = ImageDraw.Draw(shot)
d.text((60, 60), "Weekly volunteer rota", font=font(64), fill="black")
for i in range(12):
    d.text((60, 220 + i * 110), f"Row {i + 1}: table, tea or clean-up", font=font(40), fill="black")
shot.save("big.png")

def prep(path, max_side, out):
    im = Image.open(path).convert("RGB")
    scale = min(1, max_side / max(im.size))
    if scale < 1:
        im = im.resize((round(im.width * scale), round(im.height * scale)), Image.LANCZOS)
    im.save(out, quality=85)
    return im.size, os.path.getsize(out) / 1024, scale

for side in (3000, 1600, 800):
    size, kb, scale = prep("big.png", side, f"p{side}.jpg")
    text_px = 40 * scale
    note = "small" if text_px < 14 else "ok"
    print(f"max side {side}: {size[0]}x{size[1]}, {kb:.0f} KB")
    print(f"  body text now {text_px:.0f} px {note}")

crop = shot.crop((0, 0, 1500, 900))
print()
print("cropped to the part that matters:", crop.size)
print("pixels kept:", f"{crop.width * crop.height / (shot.width * shot.height):.0%}")
tiny = Image.new("RGB", (300, 200))
print("a 300 x 200 image is under 1000 px:", min(tiny.size) < 1000)
max side 3000: 3000x1800, 185 KB
  body text now 40 px ok
max side 1600: 1600x960, 69 KB
  body text now 21 px ok
max side 800: 800x480, 24 KB
  body text now 11 px small

cropped to the part that matters: (1500, 900)
pixels kept: 25%
a 300 x 200 image is under 1000 px: True

At the full 3000 pixels, the screenshot is 185 KB and the text is a comfortable 40 pixels. At 1600 pixels, the file drops to 69 KB and the text to 21 pixels, still fine. At 800 pixels, the file is only 24 KB, but the text has shrunk to 11 pixels, marked small. That is Tom’s whiteboard problem in numbers: he saved a few kilobytes and lost the writing. The cropped version keeps 25% of the pixels, but the text stays at full size, because cropping removes surroundings without shrinking what is left. That is the better trade: crop first, shrink only if you must. The last line shows a 300 by 200 image is under a thousand pixels, which is the kind of tiny picture that Anthropic’s tips warn against.

What Claude may get wrong

  • Handwriting and unusual fonts can be misread. Check any text it reads out, especially numbers and names.
  • It can miss small details, or misjudge spatial positions and counts, such as how many items are in a crowd.
  • It may describe a chart’s overall shape correctly and still misread the exact values. If values matter, give it the data.
  • It cannot know what is outside the frame, or what the image looked like before it was edited.
This is a photo of a whiteboard from a volunteer meeting. Read out all the text exactly as written, line by line. Mark any word you are not sure of with a question mark. Then turn it into a task list with the owner and the deadline for each task.

Asking for the exact text first, before any interpretation, is a good habit. You can compare it with the picture and spot misreadings before they spread into the task list.

People and privacy in pictures

Be thoughtful about photos of people. Do not upload pictures of other people without a good reason and, where it matters, their permission. Remember that screenshots can contain more than you intended: browser tabs, notifications, email addresses and other people’s names. Crop or cover first. And as with all your uploads, check the current privacy settings, which lesson 9 covered.

Common mistake: a bad picture in, a bad answer out

Sending a tiny, blurry or heavily compressed image and blaming the reply. If a person would struggle to read it, so will Claude. Retake, crop or scan again.

Scan instead of snap

For documents, a scan or a PDF export usually beats a photo, because the text is sharper and straight. If you only have a photo, a document-scanning app on your phone can straighten and sharpen it.

Photograph a page of text twice: once quickly and once straight on in good light. Upload both and ask for the text exactly. Compare the mistakes, and note what made the difference.

Practice for this lesson

Check what you learned

Five quick questions. Pick an answer and you will see straight away whether it is right and why.

Which image formats does Anthropic list as supported?

  1. Only BMP
  2. Only TIFF
  3. JPEG, PNG, GIF and WebP
  4. Only video

The upload help lists JPEG, PNG, GIF and WebP.

What happened to text when the screenshot was shrunk to 800 pixels?

  1. It became too small to read comfortably
  2. It got sharper
  3. It disappeared
  4. It changed colour

The text height fell to 11 pixels, below the 14-pixel rule of thumb.

What is better than shrinking a picture?

  1. Blur it
  2. Make it black and white
  3. Crop it first
  4. Add a border

Cropping removes the surroundings without shrinking the text you need.

Why ask Claude to read out the text exactly first?

  1. You can check for misreadings before using it
  2. It makes the reply shorter
  3. It hides the image
  4. It adds colour

Comparing the exact text with the picture catches errors early.

What should you do before uploading a screenshot?

  1. Add more tabs
  2. Cover private details
  3. Make it smaller
  4. Rename the file

Screenshots often contain names, numbers and notifications you did not mean to share.

Summary

  • Images help with text, charts, designs and on-screen problems.
  • Use large, straight, well-lit pictures and crop before shrinking.
  • Ask for the exact text first and check it.
  • Cover private details before uploading.

Next lesson

Lesson 18 covers connecting tools and staying safe.

Questions

Can Claude read handwriting?

Sometimes, but handwriting is easy to misread. Ask for the text with uncertain words marked, and check numbers and names yourself.

How big should my image be?

Anthropic's tips suggest images of around a thousand pixels or more on each side, and warn against small, low-resolution pictures. Limits vary, so check the upload help.

Can I upload photos of people?

Be careful. Avoid uploading pictures of other people without a good reason and, where it matters, their permission.

Progress is saved in this browser.