Episode 16: Tokens, tokens, who wants tokens?

Anyone who works with Claude Code will recognise themselves in my mishap.

I mean the version you run from your computer’s terminal, the one that takes you much further and lets you get your machine to do almost anything you have ever dreamed of automating.

On Wednesday, I started translating the Djtal website into English. You can rely on Google’s machine translation, or install a tool like Weglot. But if you want a quality version, one that search engines and AI assistants will favour, a real translation is worth the effort. Especially in our region, where the English-speaking community is large.

The technical framework went in without a hitch: architecture, structure, navigation logic. Nothing visible to the public at this stage. That left the heavy lifting, the content of the pages.

On Wednesday evening, just before going to bed, I put Iris on the job, as usual. She works while I sleep and in the morning it’s all done.

It’s brilliant.

Except that Iris, keen to do things properly, launched agent fan-outs.

For those who don’t know: an agent can create other agents for the duration of a task, each working on one part. At the end, the main agent collects their output, synthesises it and integrates it into the project. I use it a lot. It’s devilishly effective. But it eats tokens…

I had underestimated the scale of the work across the 90 pages of the site. And each subagent inherited Iris’s model: Fable, no less.

On Wednesday evening, my weekly Fable gauge stood at 30%, resetting on Sunday morning. Everything was running smoothly. At four in the morning, bang! Everything stopped. Limit reached.

And yet I’m on the Max 20x plan, the top tier Anthropic offers an individual user.

Nothing serious. I can switch to Opus 5.5 or use credits. But it’s a good lesson: I knew it and still got caught out.

The next morning, I went back over the project with my strategist, Sofia. From now on, on large projects, each task gets the model best suited to it.

Iris stays on Fable, which I prefer for writing. She hands review and technical work to Opus, content analysis and translations to Sonnet, and targeted research to Haiku, the most economical. Everything comes back up to Iris, who produces the final version with no loss of quality.

You always end up running short of those tokens, even on the biggest plan.

---

Laurent Cuénoud, founder of Djtal, October 2026