Building with ChatGPT

ChatGPT generates images directly inside the chat. Ask for 'an image of a quiet bookstore in autumn' and a few seconds later you have one. Free tier gets a limited number per day; Plus and Pro lift the limit.

What it's good for: rough illustrations for blog posts, mockups when you want to see a thing before describing it to a designer, social posts, kid-friendly art, conceptual sketches, mood boards.

What it's weak at: text inside images (still unreliable), specific brand consistency (each generation drifts), photorealistic accuracy of real people or places, anything requiring precise control over composition.

Image prompts work better when you describe shape and style separately. 'A reading nook by a window' is OK. 'A reading nook by a window, late afternoon light, warm color palette, photograph, soft focus on background' is better. The first part tells it what to draw; the second tells it how to render.

If the first image isn't right, iterate the same way you iterate text — describe specifically what to change. 'Same scene, but make the chair leather instead of fabric, and turn the lighting cooler.'

ChatGPT's voice mode is one of the best things about the product and one of the most under-used. You speak; it listens; it responds out loud. The response is well-paced, sounds like a person, and can be interrupted naturally if you want to change direction.

When voice is the right mode: while driving, while cooking, while walking, while you're tired and don't want to type, while you're thinking through something out loud and would benefit from a sparring partner.

When voice is the wrong mode: anything that involves reading back something visual (text, code, tables), anything you want to save or share, anything where precise wording matters and you'd lose the response when the chat closes.

Upload a spreadsheet (or CSV, PDF, text file) to ChatGPT and ask questions about it. It can summarize, find patterns, generate charts, calculate things, and re-export the result.

This used to be called 'Advanced Data Analysis' or 'Code Interpreter.' It's now mostly invisible — ChatGPT decides when to use it based on what you ask. Upload a file, then ask a question, and it'll do whatever's needed.

Honest caveat: data analysis on uploaded files works well for small-to-medium files. Very large datasets hit limits. And you should still spot-check the conclusions — the failure modes from the Mistakes piece apply to numerical analysis too.

We touched on Custom GPTs in the Skills piece. Here's the climb version: you can build your own, share it, or use thousands of public ones for specific tasks.

The use case that justifies a Custom GPT is recurring work. If you write the same kind of summary every Monday, or generate the same kind of caption every time you post, or onboard new employees with the same five questions — a Custom GPT pre-loaded with your style, your examples, and your constraints saves you the setup every time.

The use case that doesn't justify a Custom GPT is one-off tasks. Just chat normally.

Agent mode is ChatGPT trying to do things on your behalf, not just describe them. Browse the web, fill out forms, gather information across multiple pages, take multi-step actions. The vision is real; the current execution is uneven.

What works: research tasks where you'd benefit from ChatGPT visiting many pages and synthesizing, comparison shopping across sites, monitoring something that doesn't need precision.

What doesn't work well yet: anything involving login-protected sites, anything where accuracy is more important than speed, anything time-sensitive.

Honest take: if you're a beginner, you can safely skip agent mode for now. Come back to it in six months. The tooling will be much better.

ChatGPT is the broadest tool in the chat-AI category — most features under one roof. When you outgrow it, the directions to look are mostly tool-specific.

ChatGPT's climb is wider than Claude's — more features, more modes, more places to explore. The path that works for most people: chat → Custom Instructions → image and voice on demand → data analysis when you have files → Custom GPTs only when you have a specific recurring task → agent mode last, if at all.

Don't try to learn it all. Climb when you have something specific to climb toward. The features will still be there when you need them.