Help / AI workflows

Best practices for AI workflows with Cropped

Once your AI app is connected to Cropped, it can capture, read, mark up, resize and shrink images and videos on your computer. This page covers how to get great results quickly. For ready-made examples, see the workflow cookbook and the recipe library. For every tool and option, see AI tools.

#The five-step pattern

Almost every workflow follows the same five steps. Knowing them helps you ask for exactly what you want.

You don't need to name the tools. Describe the result and where it's going, and your AI picks them. Naming a step helps when you want something specific, like “number the steps” or “frame it”.

#Ask for the result, not the steps

The best requests say what the picture is for. Compare these two:

• “Crop my screenshot.”
• “Crop my last screenshot to the Settings panel, number the three switches I need to turn on, frame it, and save it as WebP for our docs.”

The second tells your AI what to keep, what to point at, how it should look and which format to use, so it gets it right the first time.

Useful things to include:

• Where it's going: a blog, docs, an issue, Slack, X, LinkedIn, a slide or an AI chat. That decides the size and format.
• What to point at: “the Save button”, “the error message”, “the three steps”.
• What to hide: “hide my email”, “blur the customer names”.
• How it should look: “frame it on the Cropped gradient”, “red box”, “numbered in violet, Ring style”.
• What to call it: “save it as settings-step-2”.

#Let your AI see before it marks

Boxes land exactly where they belong when your AI finds things first instead of guessing:

• On a Mac, ask it to find a button or label by its words (“find the Export button”). It gets a ready-made box to draw.
• Ask it to read the text when it needs everything on screen: an error message, a table, every label.
• For icons with no words, describe them (“the gear icon at the top right”) and let it look at the picture.

Read the text and Find something on screen only read pictures your AI already has, so they work with the AI capture switch off. They're Mac only for now.

#Preview before you save

Your AI can show you an edit without saving anything: ask it to “show me first” or “preview it”. It sends you the picture, you say what to change, and it saves once you're happy. This is especially useful for:

• crops around a subject, where you want to check nothing important is cut off
• several sizes at once, like a full social set
• a new recipe, before you save it for good

#Mark first, then frame and resize

Cropped measures marks on the picture as it is at that moment in the edit. A box placed before a resize stays on the button; a box placed after it uses the new picture's pixels. So when you ask for several things at once, the order that works best is:

1. Crop to the part that matters.
2. Add the marks: boxes, numbers, arrows, blur and hide.
3. Frame it, pad it or fit it to sizes.

Your AI knows this, but if you write the steps yourself, keep marks before frames, pads, fits and resizes.

#Pick the right preset

Where a picture is going decides the best format. Name a preset, or say where it's going:

Annotate an image saves with the Agent / LLM preset unless you name another one, because it's usually for your AI to look at. When an annotated screenshot is for people, ask for the Web, Share or Print preset, or use Edit an image with an annotate step, which uses your Drop Zone preset. Your own presets work too: see Optimize images.

#Save image tokens

Every picture your AI looks at costs image tokens on your AI plan. Cropped keeps that low by default, and you can help:

• Pictures go back to the AI as a small preview with the Agent / LLM preset. That's enough to see layout, buttons and most text.
• Ask for the full-size picture only when the AI needs to read small text. It's sent up to 3.7 MB and 8000 pixels on a side.
• When the AI only needs to save a file, it can ask for no picture at all, just the details and the path.
• Read the text returns words instead of a picture, which is often cheaper than looking.
• For videos, Look at a video sends 6 small stills by default. Ask for a closer look at one part rather than more stills of the whole thing.

#Turn good edits into recipes

When an edit comes out right, say “save that as a recipe called Blog hero”. A recipe keeps the steps, the sizes to make and the format or preset. Then any AI app connected to Cropped can run it by name:

• “Run Blog hero on my last screenshot.”
• “Run Launch set on hero.png.”

Recipes also show up in your AI app's prompt menu as Recipe: its name. A recipe runs first, then any extra steps you ask for. If you'll add marks to a picture, add them before running a recipe that frames or resizes: “annotate it with the Print preset, then run Docs frame on the result.” See the recipe library for 18 recipes to start from.

#Work through folders

AI apps that can see your files, like Claude Code, Cursor and Codex, can run Cropped on every file in a folder:

• “Shrink every PNG in ./docs/images with the Web preset.”
• “Run Blog hero on every JPG in ~/Desktop/photos.”

Claude Desktop works with the files and captures you name, unless you've also given it a tool that lists folders. For very large batches, Cropped's own watch folder or the Studio can be quicker.

#Videos: let your AI look first

Your AI can't watch a video in real time. It sees stills with their times, so the best video requests let it look first:

1. “Look at my latest recording and tell me where I open Settings.”
2. “Keep from there to where I click Export, without the sound.”

Your AI sees pictures, not sound, so it can't find a spoken word. Tell it the time instead (“the part at about 0:40”). Keep GIFs to a few seconds: they grow fast. For anything longer, ask for an MP4 or WebM.

#Keep private things private

• Ask your AI to hide emails, keys, tokens, phone numbers and names before sharing. Hide is a solid block that can't be undone in the saved file.
• Use blur for faces and things that should look softened rather than blacked out. For secrets, hide is the safer choice.
• Use the Redact before sharing prompt to have your AI find private details first and ask you before hiding them.
• The AI only sees what you let it: a screenshot you take, a capture you name, or a window it captures with Let AI apps capture without asking on. Every capture without a click flashes and plays the capture sound.

See Privacy for exactly what goes to your AI service.

#Work alongside your other apps

Cropped saves real files on your computer, so it works alongside the other tools your AI app has:

• Code tools: in Claude Code, Cursor or Codex, your AI can capture the app it's building, check the result and save screenshots into your repo for a README or pull request.
• Image makers: an image made by another AI tool can be finished in Cropped. Edit an image takes a file, an https link or the image itself, then fits it to sizes and saves it with your presets.
• Voice tools: an audio file from a text-to-speech or recording tool can be the voiceover in Edit a video.
• Issue trackers and chat: after Cropped saves a file, your AI can attach it wherever its other tools allow, using the saved path.

#When something goes wrong

• A box is in the wrong place: ask your AI to find the element first, or to read the text, and try again.
• A crop cut off the subject: ask for a preview and say what to keep, like “keep the whole dialog”.
• The picture came out smaller than the size you asked for: the original was smaller. Ask your AI to enlarge it, or start from a bigger picture.
• Your AI asks you to click a window: turn on Let AI apps capture without asking in Preferences › App › AI tools (Mac).

More in AI tools troubleshooting.

Still stuck? Click the ? button at the bottom left of the Cropped Studio to ask, send us a message, or request a feature.