Context management is the skill. The tools are interchangeable.
Most operators frame their AI question wrong. They ask which model is best. I stopped asking that about a year ago. The better question is: which capability does this task actually need, and which tool clears that bar at the lowest cost?
Here is my current setup, and why I built it this way.
Coding
I have never written code in my life. That means I cannot optimise for speed. I optimise for not breaking things.
Before a single line gets generated, I research. YouTube, Reddit, technical forums. I need to understand what I am building and which tools fit it. That research goes into a single brief. I debate the brief with Claude first to spot architectural problems. Once the brief survives that, I build. I use Claude Code as the primary builder and am currently testing Gemini’s IDE for extended context. My rule is simple: if a tool fails the same task twice, I replace it.
Pressure-testing arguments
I do not use AI to get answers. I use it to break my reasoning.
The prompt I run most often is not “explain this to me.” It is: “where is the logic gap?” or “what assumption is false?” I run the same argument through Gemini, ChatGPT, and Claude. If a line of reasoning survives three separate adversarial passes, I treat it as ready for the real world. I tell each model explicitly: do not be polite. That single instruction changes the quality of the output more than any other variable.
I start in Gemini because its long context window handles a large brain dump without compression. ChatGPT and Claude run the critique.
Drafting
AI does not write my final drafts. It edits them.
I load the goal, the audience, and the structure. I apply strict style rules upfront to prevent generic output. Then I run four to five rounds of edits with AI. When I mark something final, I walk away. I read it again the next morning. Errors always appear with fresh eyes. Only after that pass do I use AI for a final clarity check.
Research
When I need market data, technical documentation, or sourcing, Gemini wins. Gemini’s Google Search connection reduces the friction of opening twenty tabs. But no model builds a complete report on its own. The first output is always shallow.
My process: let AI build the base. Read it. Find the gaps. Build a targeted question list for what is missing. Investigate those specific areas manually or with focused prompts.
The part that actually matters
The specific models above account for roughly ten percent of my results. The other ninety percent comes from how I manage context across sessions.
Most people rebuild their AI context from zero every time. They re-explain their business, their tone, their constraints. That is waste. It also means the model is always working with an incomplete picture of what you are trying to do.
Every project I run has one master document in Google Docs. It holds the current state, the goal, and the constraints. I update it continuously. Every AI session starts by reading that document. The model always sees the latest version of reality. I never repeat myself. I swap models without losing momentum.
Tools are commodities. The system you use them in is not.