Tipping Point

Ever since ChatGPT arrived on the scene, complaints that it failed this or that task were met with the reply, "You must have used the wrong prompts." But in the early days, more often than not, that was disingenuous. Many tasks were simply beyond its reach, no matter how clever you'd prompt.

My recent experience using Claude Code for real development suggests that we have reached a tipping point. You still have to meet the AI halfway by providing good prompts, or more precisely, good context via the various project-wide instructions, extra skills, MCP connections etc. But at least now, that effort gets rewarded with surprisingly good performance.

Compare it to a situation in which a manager delegates a task to a team member and gets a poor result. The fault is typically shared between the manager and the team member. Did the manager ask for the impossible, explain the task poorly, or provide insufficient context and guidance? Or is the team member just not capable?

A year ago, if you told me you got poor results using AI for coding, I'd have told you: "It's not you, it's them." At this point (Claude Code Opus 4.5, ChatGPT 5.2, Gemini 3), we're sliding closer to "It's you. Get better at using these tools." So now I'm running an experiment for the next little while: Can I get Claude Code to write all my code? And if not, what can I do better in how I've set it up? I'll assume a poor outcome is 100% on me. That assumption won't always be true, but it will be useful.

I sense that we'll see similar leaps and crossings of tipping points in other domains as well, where we suddenly go from "meh results and even that only if you prompt really well" to "here's how to set it up to get amazing results."

If you have been disappointed by AI results recently, give it another go with the latest models. You might just be pleasantly surprised!