Vision

Vision Enabled Models, for explaining images, doing OCR, extracting Text, and other vision tasks.

Vision

Models for OCR, multimodal reasoning, and image understanding.

Warning: these model recommendations are personal opinion, reflecting individual preferences and tradeoffs rather than a universal ranking.

Prompting

Vision models work best when you describe the image task precisely and define what kind of output you need. The image itself is only part of the prompt; the instructions around it matter just as much.

1. State the visual task explicitly. Say whether you want OCR, object identification, chart reading, screenshot debugging, or a natural-language description.
2. Tell the model what to focus on. Mention the relevant region, objects, text, or relationship in the image.
3. Define the output format. Ask for JSON, bullet points, extracted text only, a confidence note, or a short explanation depending on your use case.

Below are examples of well-structured prompts, followed by examples of what to avoid.

Effective Examples

OCR With Structure

This prompt asks for targeted extraction and constrains the output.

Extract the text from this receipt image.

Output requirements:
- Return valid JSON.
- Include the keys `merchant`, `date`, `total`, and `line_items`.
- `line_items` should be an array of objects with `name` and `price`.
- If a field is unreadable, use `null` instead of guessing.

Why this works:

• Specifies the task type: The model knows this is OCR and structured extraction, not summarization.
• Defines the schema: The requested JSON shape reduces ambiguity.
• Prevents hallucination: Using null for unreadable fields is safer than filling gaps.

Screenshot Analysis

This prompt narrows the model's attention to a concrete UI problem.

Analyze this screenshot of my web app and identify the likely layout bug.

Focus only on:
- the header,
- the first card in the grid,
- and any overlapping elements near the right edge.

Answer in 3 parts:
1. what looks wrong,
2. the most likely CSS cause,
3. one specific fix to try first.

Why this works:

• Narrows the inspection area: The model does not need to comment on the entire screenshot.
• Requests a diagnostic structure: The three-part answer keeps the result practical.
• Targets actionability: Asking for one specific fix first reduces noisy speculation.

Poor Examples (What to Avoid)

The following prompts are less effective because they leave the model guessing what to inspect and how to respond.

Unfocused Image Prompt

This prompt does not define the job.

What do you see here?

Why it's ineffective:

• No task framing: The model does not know whether the user wants OCR, description, diagnosis, or extraction.
• No output target: The response could be verbose, generic, or irrelevant.
• No attention guidance: Important details may be ignored.

Guess-Heavy OCR Request

This prompt encourages unreliable extraction.

Read this blurry document and fill in whatever is missing.

Why it's ineffective:

• Invites hallucination: "fill in whatever is missing" tells the model to invent unreadable content.
• No schema or format: The user cannot easily validate or reuse the result.
• No quality threshold: The model is not told when to admit uncertainty.