HeyGen, Claude, ChatGPT, and Cursor: Four Tools, Four Jobs

I use four AI tools regularly. They do four different jobs. The habit that took me a year to build is knowing which one to open before I open it.

This is my personal stack: side projects, experiments, and community work. Not a prescription for anyone else's environment.

ChatGPT. Conceptual design and images.

For: logos, image generation, layout concepts, quick research, loose outlines, playing devil's advocate.

I open ChatGPT when I need to move across unfamiliar terrain without committing yet. Random rabbit holes. A rough structure for something I have not written. Someone to argue the other side before I lock in a direction.

Where it really earns its tab: visual work at the concept stage. Logos. Any image generation. Layout ideas when I am still exploring what the thing should feel like. Chat is best when the goal is conceptual: fast drafts, variations, and directions I can react to before I lock anything in.

What it is not: my templating or compositing surface. When I need to layer images together or hold a design to a repeatable structure, I go to Claude.

Claude. Documents, presentations, and design systems.

For: documents and decks that need to hold together, templated layouts, layering and compositing images from ChatGPT, longer reasoning when the work needs to stay coherent.

Claude is where writing and design work graduate from concept to structure. I take rough directions from Chat and turn them into documents I can share: memos, briefs, outlines that survive a second read. Same for presentations: deck frames, slide systems, narrative arcs I can run variations through without reinventing each slide from scratch.

For visual work, Claude is where Chat-born images get layered and composed: combining assets, refining how elements sit together, building layouts that need to repeat or scale. Chat is great for the conceptual pass. Claude is great when the output needs a template: a page system, a deck frame, a layout I can hand off or iterate on without starting from scratch each time.

Claude Cowork is a separate tab in my stack for agentic work: MCP connectors, custom agents, and skills when the job is orchestration across tools rather than a single long chat. Wiring apps together. Running repeatable workflows. Letting an assistant act on my behalf instead of only suggesting the next paragraph.

I started building websites with Claude before I moved to Cursor for sustained development. That split still holds: Chat for how something should look and feel at first; Claude for how the words and layouts hold together as a system; Cowork when I want agents and connectors in the loop.

Cursor. Building and vibe coding.

For: ship software, refactor, debug, learn by doing.

Cursor is not a writing tool. It is a build surface.

This is where I built this personal site. Where I spin up experimental websites to test ideas before I know if they deserve a real name. Where the stack stopped being vocabulary and became muscle memory.

I started with Claude for web projects. Cursor won for ongoing development: more affordable for the volume of building I do, and purpose-built for working inside the codebase instead of guessing from pasted snippets. Claude Cowork handles MCP and agents on the orchestration side; Cursor handles MCP inside the repo while I vibe code.

If ChatGPT helps me explore and imagine, Claude helps me write and template, and Cowork helps me wire agents across tools, Cursor lets me exist in the artifact.

HeyGen. Video and voice experiments.

For: avatar-led video, testing how a message lands when spoken, prototyping broadcast-style content.

HeyGen earns its place when the deliverable is motion plus narrative. I am mostly experimenting right now: finding the best-quality clone of myself for a weekly AI-generated local news segment I want to run for my community.

The point is not to replace being on camera. It is to lower the cost of hearing a story out loud before I commit to how it should be told.

How the stack sorted itself

The evolution was not "better models." It was clearer jobs.

Critique as you build

Across all four tools, the habit I lean on most is turning AI into a critic while the work is still small. Halfway through a layout, a draft, or a flow, I ask where the weak spots are: what I skipped, what assumes too much, what breaks if someone is tired or skeptical. Chat is good for devil's advocate. Claude is good for stress-testing a deck or document before I share it. Cursor catches implementation gaps before they compound. The point is not pessimism. It is finding the hole while the fix is still cheap.

The rule I keep

If I cannot name the job, I do not open the tool.

That single habit saved me from the worst version of AI work: busy, shiny, and strategically empty.