Stop having a favourite. ChatGPT, Claude and Gemini become a bench, and you learn to make the model write its own prompts.
You would not use one lens for every shot. These models have different shapes, and the differences are stable enough to plan around even as versions change. ChatGPT is the generalist with the widest tool surface: images, code execution, file handling, custom GPTs, voice. Claude is the long form and code instrument, it holds a large document in view without losing the thread and follows a complicated brief closely. Gemini sits inside Google, next to your Docs, Sheets, Gmail and Drive, with genuinely strong handling of very large inputs.
These are tendencies, not laws, and they move with every release. The skill is not memorising which is best. The skill is testing all three on your own real task and forming your own view, which is exactly what today asks of you.
ChatGPT when you are not sure what the task needs, start here. Claude when the task is read all of this and be careful. Gemini when the material already lives in Google, do not export it somewhere else.
Every one of these tools has a place to write standing instructions that apply to every conversation: custom instructions, project instructions, a saved style. Most users never open it. This is the highest leverage ten minutes in the course. Written once, it stops you re-explaining yourself forever.
Write it as constraints, not compliments. Be concise is weak. Answers under 200 words unless I ask for depth, no bullet lists unless I ask, no restating my question back to me, that is enforceable.
Meta prompting is asking the model to build the prompt. You are not a prompt engineer yet, on Day 2. So delegate the first draft: before you write anything, ask me the questions you need answered to do this well, then write the prompt you would want to receive.
Three things happen. It surfaces decisions you had not made. It produces a stronger prompt than you would have written. And you learn prompt structure by reading good examples rather than by being lectured. Variants worth keeping: interrogate first, critique the brief, and rebuild my prompt so a different model would produce the same result.
A good result is usually turn four, not turn one. Say what is wrong. Ask for three more options. Ask it to keep paragraph two and redo the rest. Ask why it made a choice, then overrule it.
Corollary: when a thread has gone bad, do not fight it. The mistakes are in the context now and they pull everything after them. Start fresh with a better opening prompt. Restarting is not failure, it is cheaper than salvage.
All three tools now separate the chat from the thing being made. Use it, iterating on a persistent object beats scrolling back through chat to find version three. Projects, where available, are how you stop repeating context: load the brand guide, the tone rules and the references once, then every conversation inside inherits them.
NOVA reacts, nothing is scored, nothing is stored against you.
Build the Pong game. Ask Claude for a single self contained HTML file with a playable Pong game, keyboard controls, a score, and a small particle burst when the ball hits a paddle. Save it, open it in your browser, and when something breaks, paste the error back and ask for a fix. You have just written and debugged software without writing code.
Run the same substantial task through all three models, then use a fourth conversation to judge the three outputs against your criteria. Keep the scorecard. It becomes your routing rule.
Day 2 in progress
Tomorrow, Day 3: prompting stops being typing and starts being directing. Six components, few shot, chain of thought.