More frugal with AI

In the early days of 2023 and 2024, there were many strategies to get better performance from an AI using less tokens. The idea has been popular for a long time now. It varied from offering $1 billion dollars in a prompt to the famous “Make no mistakes” meme. Now, our question is: Are the same strategies we used 2 – 3years ago still a good idea today? That is an important question as new AI models improve on reasoning/thinking and show improved horizon intelligence seeing the bigger picture.

If we look at benchmarks that are difficult to optimize for such as Minebench, we see a variety of AI models that are capable of elaborate constructions that look very complete. Just a year or two ago, these shapes would be simple and vague interpretations at best. That has rapidly changed much like “Will Smith Eats Spaghetti” has proven for other types of GenAI.


Skills, plugins, and more?

OpenAI’s ChatGPT/Codex and Anthropic’s Claude Code are sometimes used with elaborate configurations of extra software such as special instructions. Skills like the popular Superpowers skills bundle claim to help AI models act more like human developers following the same types of processes to mimic the results. While popular with some, others claim that these aren’t needed with the latest most powerful AI models wasting tokens instead. If your project doesn’t need a security review cycle, review gates, unit tests, and more, you might be better off with a lighter method.


“Make no mistakes”?

The ever popular meme of “Make no mistakes” comes to mind often and it isn’t far from the truth. A recent piece of research suggests using “Do not guess” to cut made up claims from 70% to 20%. This follows other similar guiding documents such as those from Andrej Karpathy in CLAUDE.md

The “If something is unclear, stop. Name what’s confusing. Ask.” and similar instructions help prevent mistakes before they get baked into code, specs, and the final product. This helps save time and tokens keeping your projects moving forward. These can also be helpful for complex ChatGPT Work projects and aren’t solely applicable to Codex or Claude Code alone.

We hope this helps you maximize your tokens. Stay tuned for more to help you work with WordPress, AI, and more!