Getting Better with AI

I'd say it'd be wise to stop chasing the newest tools. It's not going to come without a loss.

I see people jump from tool to tool expecting the next one to finally be the thing. It doesn't work like that. Each tool has a shape. Claude Code is great at long-running tasks. Codex on GPT 5.5 is great at small, specific things. If you're on Opus 4.7 and you hop over to Codex 5.5 expecting an upgrade, you're going to be disappointed. You're not switching to a better tool — you're switching to a different one, and your needs aren't going to be met the same way.

Set higher expectations for yourself instead. Pick one. Learn it. I know that's counterintuitive, but if you watch how the market actually moves, there's a pattern on roughly a 3-month cycle.

The 3-Month Cycle

OpenAI was so dominant that even my dad — a janitor who only speaks some English — was using it. Then Anthropic started taking market share. They stopped looking like the brown-paper-bag version of AI and became the darling of the industry. Then they bricked their own product and started making anti-consumer decisions right as Open Claw came out. OpenAI shipped 5.5, opened up OAuth to third-party harnesses, and shifted the narrative. The YouTube-verse pivoted overnight to pitching Codex as the best model. Now it's going to be Cursor with V3 and the xAI partnership. Oh, and for some reason Anthropic is in with xAI too. I guess Elon Musk hates Sam Altman more than he hated Dario and the church of Claude.

It's a constant cycle of shifting perception, and chasing it is a waste of your time. Pick a tool. Master it. Build your own harnesses and tooling around it. Explore other tools — don't ignore them — but watch how much time you sink into onboarding. That cost usually comes out of your pocket in money, time, or both.

So How Do You Actually Apply Your Tooling?

Say I'm a business manager who needs to analyze data. The goal isn't just to get the task done — the goal is for me to get smarter. AI will produce stuff for you, sure, but do you actually understand what it gave you?

You need to think like a product manager building a tool for customers, except you are the customer. Can you extend what you build? Can you add features without degrading quality? Can you keep building on the same thing for days or weeks without breaking what you treasure? Can you confidently hand it to a coworker without it blowing up their system?

The Workflow I'd Actually Run

1. Ask: “What are 10 techniques data analysts and scientists use to analyze data?”
2. Have it summarize each one into a 3-line blurb I can actually understand.
3. Ask for a 5th-grade-level example of each.
4. Then a 10th-grade-level example.
5. Then an example using the type of data I actually have.
6. Give the AI a real sample of my data.
7. Now here's where most people miss the barn they're swinging at — what's your eval criteria? How do you know the answer is right for you? It might be technically correct but useless for your context.
8. Context — does your AI have enough context about your problem to even produce a correct answer? This, like other terms at the moment, is being hijacked and overused. If you want to explain something to someone you give them context, info, the history, tips, the drama, gossip, the whatevers… etc. That's context.
9. Steering docs — you have to write documentation FOR your AI. It needs a reference base it can pull from.
10. Label the cases where it was right.
11. Label the cases where it was wrong.
12. What areas does it need to research to understand your world well enough to be relevant?

A Few Quick Ones People Argue About

• Is prompting important? Yes.
• Is prompting everything? No.
• Will prompting alone get you the best results? No.
• Is RAG relevant? Yes.
• Can you implement it yourself? Kind of. You need to understand why every major AI tool ships with a Projects feature. That's where you upload your supporting documents — it's a layperson's RAG. Use it.

The solution is easy. A little bit of effort into curating your data and you're gucci. Why do you think so many companies are NOW finally caring about data engineering and data quality?