There's a phase every creative professional goes through when they start exploring AI tools. I call it the shiny tool phase. You hear about something new, you sign up, you get excited, you imagine how it's going to change everything. Then two weeks later you haven't opened it, and something newer has already caught your attention.
I've been through that phase. I'm mostly out the other side now. What's left is a surprisingly small toolkit that I use every day, surrounded by the wreckage of subscriptions I cancelled, local setups I stopped maintaining, and platforms I'll probably never log into again.
This isn't a "best AI tools" list. It's an honest look at what the experimentation phase actually costs, in time and money, and how I eventually stopped chasing and started working.
The Tools That Survived
Claude (Anthropic). This is my primary AI tool now, full stop. I use it for writing assistance, project planning, troubleshooting, research, and conversation. The Pro subscription ($100/month) gives me access to Claude Code, which has become the backbone of how I build software. I also use Cowork for managing tasks and workflows across projects. Claude is where I do the thinking.
What won me over wasn't a single feature. It was the way Claude handles complex, multi-step work. I had already developed a task-based system for building software, where every update creates a task document I can review later if something breaks. Claude Code turned out to be exceptionally good at managing that system. It understood the workflow I had already built and made it better.
ChatGPT (OpenAI). I keep my $20/month subscription for two specific reasons: image generation and the voice feature. When I need a quick image, ChatGPT handles it well. And I genuinely like being able to listen to responses rather than reading them. It sounds like a small thing, but when you're working long hours at a screen, switching to audio makes a real difference. Claude doesn't offer this in its app yet, and that gap keeps ChatGPT in my rotation.
Google Gemini. I use this occasionally, partly for the same reason I keep ChatGPT around (it also has voice), and partly because its image tool, Nano Banana, does something genuinely different. It lets you doodle edits directly onto an image and have the AI interpret your intent. It's a surprisingly intuitive approach to image editing that I haven't seen replicated elsewhere.
Cursor and Claude Code. Here's an honest admission: I pay for Cursor, but I barely use its built-in AI features anymore. Almost everything I do runs through Claude Code. Cursor has become more of an editor that happens to be open while Claude Code does the heavy lifting. I've also used Antigravity in a similar way. At some point I need to evaluate whether the IDE subscription is still worth it, or whether I'm just paying for a comfortable text editor.
The Tools That Didn't Make It
GitHub Copilot. I had the yearly subscription early on. It was fine. Competent autocomplete, occasionally helpful suggestions. But I never leaned into it. When I started using Cursor's standard tools, Copilot felt redundant, and I let the subscription lapse. It wasn't bad. It just never became essential.
Midjourney. I tried this when image generation was the hot topic, and the output quality was impressive. But the Discord-based interface was miserable. Having to manage prompts and outputs inside a chat server felt hostile to any kind of creative workflow. I know they've launched a proper website since then. I've never gone back to find out if it's better.
OpenAI Codex. Works well for the same task-based development system I use with Claude Code. But with my Claude Pro account, I can work continuously without worrying about separate usage limits or costs. Codex became unnecessary once Claude Code proved it could handle everything I needed.
The Local Setups I Built and Walked Away From
This is the part that's harder to talk about, because it involves real time investment.
I have Comfy-UI installed and running on two Windows machines. I've created custom LoRAs. I have Ollama set up on one machine and LM Studio on another. At one point, I was genuinely excited about running AI locally, having full control, no subscription costs, no rate limits.
Then I switched to Mac.
It wasn't planned as an AI decision. I'd been a PC user for decades, but after two Intel processors died on me, friends recommended trying Mac for development. I started with a Mac Air, liked it, and invested in a Mac Studio for heavier work. And something shifted. The Mac felt clean, focused, purposeful. I never felt right cluttering it up with local AI infrastructure the way I had with my Windows machines.
Meanwhile, the cloud tools kept getting better. Claude got smarter. ChatGPT added features. The gap between what I could run locally and what I could access online shrank to the point where maintaining local setups felt like maintaining a hobby, not a workflow.
I still use those Windows machines over the network. They're not gathering dust. But the frequency has dropped off significantly as the cloud tools have closed the gap.
The Tool I Built Because Nothing Fit
When I couldn't find an AI writing assistant that matched my process for long-form creative writing, I built one. DraftDesk.io is a tool I created specifically for my own workflow with screenplays, fiction, nonfiction, and series work. It runs through my local Claude account rather than paid API calls, because I'm deliberate about not letting costs spiral.
This is the part that matters for other creative professionals reading this. Sometimes the right tool doesn't exist yet. Sometimes the best use of AI isn't finding the perfect app. It's using AI to build the app you actually need. DraftDesk.io isn't a product I'm selling. It's a solution to a problem I couldn't solve with anything on the market.
What I Actually Spend
I try to be transparent about this because the costs add up faster than people expect.
My current monthly spend is roughly $120. That's Claude Pro at $100 and ChatGPT Plus at $20. I have Cursor on a yearly subscription too. I also have small amounts of API credits scattered across various platforms, but I'm extremely cautious about spending that way. The DraftDesk.io decision to use my local Claude account instead of API calls was specifically to keep costs predictable.
The stuff I'm not paying for anymore: Copilot (yearly, cancelled), Midjourney (dropped after the trial period), and the electricity and maintenance overhead of running local AI setups that I've since abandoned.
What I Learned
The shiny tool phase is expensive, but it's probably necessary. You can't know what works until you've tried what doesn't. The mistake isn't experimenting. The mistake is not being honest with yourself when something isn't working and continuing to pay for it anyway.
Here's what I'd tell any creative professional just starting to explore AI tools:
Start with one general-purpose AI (Claude, ChatGPT, or Gemini) and actually use it for a month before adding anything else. Pay attention to what you reach for naturally and what you have to remind yourself to open. The tools that stick are the ones you forget are "AI tools" because they've just become part of how you work.
And be honest about your budget. A $20/month subscription that you use daily is infinitely more valuable than five $20/month subscriptions where you're "going to get around to" using most of them. The AI tool landscape will keep churning out new options. Your job isn't to try them all. Your job is to find the ones that let you do your actual work, and then get back to doing it.