by Luis Rodrigues
What I’m building with AI + one hot take + five links worth your time.
The Take
This week I opened Pydantic's Logfire to see how thoughtled.ai chatbot is performing between 2 providers. Same prompt and tasks, and I found exactly what I expected: the ranking changed.
GPT now uses far fewer tokens per task and, since OpenAI's price cuts, is also cheaper.
This is something people building using AI need to consider. Every month there are new versions, sometimes with very different tokenizers, that completely change the scoreboard.
This week Microsoft decided to do something about the same problem. Jay Parikh, executive vice president, sent a memo setting GPT-5.6 Sol (OpenAI's priciest model) as the default for internal use of GitHub Copilot.
"We are not optimizing for fewer tokens; we are optimizing for more impact per token."
If you're thinking this is cost control, you're missing the full picture. Microsoft is crazy profitable, and they beat Wall Street forecasts. So this is a company that realized that tokenmaxxing was never the strategy.
Having a default makes sense. Now, the metric I'm not convinced of. Budget tells you what a division spent, never whether the number moved because someone was wasteful or because a new model version changed the math.
Per token is the wrong denominator: we should all start focusing on finished tasks.
If you're building, price one real task end-to-end across multiple providers this week (including the Chinese) and see which one is actually cheaper.
If you're getting ready to start, here's a habit: scoreboard changes every month; always re-measure.
The Build Log
Someone sent me an email this week following up on the Claude workshop because he had copied some Claude configuration online and Claude was slower.
He just wanted to see my .claude folder structure. So here it is:
Every file in this folder has a use, and you should know how they all work to avoid wasting precious tokens.
Part of this folder is loaded in all sessions before you even write a single word.
CLAUDE.md is loaded in all chats. So are the Rules if you don't scope them to file paths. So if you have 500 lines in there, it's 500 lines of context that you're paying all the time.
And the worst part of paying in context tokens is that the more data you have in there, the worse it is for the model to decide what is really important.
Skills/commands are loaded only when they're needed for the task at hand, or when you explicitly ask to load them.
Agents run on their own context window and only return the result. So, for example, a security review takes only a few lines in your main session, compared to the dozens of files that the main agent chat would load to complete it without agents.
settings.json costs nothing in tokens, but it's one of the most important files. It doesn't have instructions; it has conditions: what the model can and can't do, what files it can't touch, or which scripts run around a certain action.
Hooks are scripts that are executed when certain conditions happen, for example, before a tool call, after an edit, and Claude can't avoid running them
If you're working with more people and using git, the correct way to proceed is to commit all files except those with .local.
My settings.local.json and CLAUDE.local.md are always listed in gitignore to ensure I don't make any mistakes.
So in thoughtled.ai, the CLAUDE.md is short and boring (like the stack). Just has basic conventions and stack information. If Claude needs info for a given part of the code, it can load it from the CLAUDE.md in the directory or via the information in rules
If you're building, make sure you split your main CLAUDE.md file in multiple. Create one per directory so it's only loaded when Claude reads those files.
If you're not building yet, keep this email in mind; it will come in handy later.
On My Radar
A prompt with about 25 tokens (16 characters) sends more than 9,000 tokens. Another proof that harnesses are getting bigger and more expensive.
A common format for packaging Agent Skills and MCP servers into distributable plugins. Claude Code not in the supported list.
Alibaba's 2.4T-parameter flagship went GA at $2/$6 per million tokens. Should have a large improvement in coding and agentic work.
A 3B open-weights safety classifier under Apache 2.0 that takes your moderation policy as a plain-language question for evaluations and runs on one 16GB GPU.
Spending caps for agents that will be enforced by the infrastructure rather than by a line in the system prompt.
Need another pair of eyes on your AI project?
One consulting slot available for August → Book a call
Thanks for reading.
See you next week. Go build something.
Luis
Know someone who'd enjoy this? Forward it.
New here? Someone forwarded this to you? → Subscribe to Build What Matters: https://luisrodrigues.ai/
