For more than 20 of my 27 years running commercial operations, I was never once in a position to build the tools I needed. I could see exactly what would help. A lead-scoring flow, a reporting job that ran itself, a small system that chased the thing everyone forgot to chase. And every time, I wrote a brief, joined a queue, and waited. On internal IT. On a data scientist who had 40 other requests ahead of mine. On an agency that would quote six figures and come back six months later, by which point the brief had moved and the thing they delivered was close to what I now needed rather than exactly it.
The idea for the tool I shipped last month arrived on a Saturday run, somewhere around kilometre 15. That is usually where they arrive. By the following weekend I had a working pilot. Two weekends after the run, it was live. I built it on evenings and weekends, around a full-time job and training twice a day. And I still cannot read code.
I want to be precise about what changed. Who gets to hold the tools changed, and how you hold them. So this is the operator's version of a tool review: the way a commercial person who ships real products actually splits the work across the stack, where each tool earns its place, and where each one fails me.
The stack, by the job it does
I run four tools, and I run them for different reasons. The split matters more than any single one of them.
Claude Code is my daily build partner. This is where the products get made. I run it inside VS Code rather than the desktop app, and the reason is the whole point: in VS Code I can see the files. I watch the folders fill up, I read the changes as they land, I see the difference between what the code did yesterday and what it does now. Claude edits those files itself. I do not hand-change them, because I could not. I read, I question, and I instruct, the way I have briefed and corrected analysts for 27 years.
The desktop app is a smoother conversation, but it hides the files behind the chat, and once you have worked where you can see everything, working blind feels like managing a team through a closed door. I am leaning toward dropping the desktop app entirely.
Claude chat on my phone is for quick questions. A definition, a fast sanity check, something I want answered while I am walking. Nothing gets built there. It is the tool I reach for when I have a hand free and a small question.
ChatGPT on the laptop is my thinking and writing partner. When I need to reason something out, work through a plan, pressure-test an idea before I commit a weekend to it, that is where I go. It writes genuinely well, well above what I expect from these tools, and its images from a single prompt are the best I have used, realistic and finished in one pass. I have run it hard and never once hit a usage limit. When I need to sit with an idea rather than build it, ChatGPT is the room I think in.
Codex runs my autonomous agents. It runs on GPT-5.5 at high effort, and the agents I have built on it hold up: a lead-finding agent, content generation, a WhatsApp automation that runs on its own. It is newer to me than Claude Code, and I trust it for the jobs that run without me watching.
That is the map. Claude Code builds, ChatGPT thinks and writes and makes images, Codex runs the agents that run themselves, the phone catches the small questions. Four tools, four jobs, and I move between them without much thought now because I have done it enough times that the boundaries are obvious.
Casting the model to the task
Inside Claude, I do not use one model for everything. I cast the model to the task, the way I used to cast people onto a cross-functional team. Some people you put in charge of holding the whole thing together and catching the errors. Some you give the hardest, most delicate piece of the build. Some you hand the strategy and the specification, and some you give the routine work that just needs to get done reliably. You would never staff all four of those roles with the same person. The models are the same.
So Fable 5 is my orchestrator and my quality control. It is the one I trust to hold the whole build in its head and catch what is wrong, and it is also where I send the hardest, most delicate coding. Opus, run at high effort, is where I go for strategy sessions and for building the backbone, the spec that Fable 5 then executes against. Sonnet handles the routine work inside the agents, the tasks that repeat and do not need the heavyweight. On ChatGPT and Codex I run GPT-5.5 at high effort for the reasoning-heavy work.
It is easy to pick one tool, or one model inside one tool, and stay there. You find a setting that works once and it is comfortable to leave it, because moving between them takes attention, and we are all busy. What has worked better for me is treating the stack as a stable of specialists I cast by job. The casting is where the judgment sits: reading a task, sizing what it actually demands, and matching a model to it the way you would decide who to staff on a piece of work.
I did not learn that from a tutorial. It came out of using these things every day, watching the tasks get more complex, and being curious enough to notice which model was good at what.
Where each one fails me
None of this is frictionless, and I would not trust a tool review that pretended otherwise.
Claude's usage window is the real one. I hit the ceiling and sometimes wait several hours before I can keep building, which on a weekend evening is expensive when the evening is the only build time I have. I have already moved from Pro to Max, and I expect to need the top tier soon, which runs around 100 CHF a month. That is real money, and I am telling you the number because the "you can build for almost nothing now" story quietly skips it.
There is a second discipline Claude forces on me, and it took a while to learn. I have to manage the context window actively. The signal is quality: the moment Claude's answers start getting worse, drifting, missing things it caught an hour ago, that is not a reason to push harder. That is the window telling me it is full. So I run a wrap-up, save the state, and start a fresh window. Learning to read that drop as a signal rather than a frustration changed how much I get done in an evening.
ChatGPT does not read or write the files on my laptop, which is exactly why it stays my thinking partner and not my build partner.
Codex I am still growing into. It is not yet as heavy a workhorse for me as Claude Code, partly because I have simply spent fewer hours in it. That is on me, not the tool.
Why I am not writing you a how-to guide
At the pace these tools move, a how-to guide goes stale almost the week it is written. I spent a good while hunting for that guide myself, the definitive one. The best prompt library, the perfect stack, the one workflow that settles it. The pull is completely understandable, and I still feel it. The model names change. A tool I called a weakness ships the feature next month. A fixed guide is a photograph of a thing that will not hold still.
Do the build yourself, even slowly, even badly at first. Keep experimenting. Ask ChatGPT or Claude to help you find the better solution, and then push on what they give you. I learned the hard way to do that, because they are confidently wrong often enough that I stopped taking the first answer at face value. The models are analysts who are fast and sometimes flatly mistaken, and I have found it pays to check them. The skill is the experimenting. Everything else has a shelf life measured in weeks.
The two weeks from a Saturday run to a live product is a story about a commercial operator who spent 20 years waiting in a queue, and who does not wait anymore. What replaced the wait is smaller than it sounds: a willingness to sit down on a weekend evening, open the files, cast the right model to the job, and go back and forth with it until the thing works. Most weekends I am still figuring it out as I go.