The magic is in that codify step. LLMs are stateless. If they re-introduce a dependency you explicitly removed yesterday, they'll do it again tomorrow unless you tell them not to. The most common way to close that loop is updating your CLAUDE.md (or equivalent rules file) so the lesson is baked into every future session. A word of caution: the instinct to codify everything into your rules file can backfire (too many instructions is as good as none). The better move is to create a setting where the LLM can easily discover useful context on its own, for example by maintaining an up-to-date docs/ folder (more on this in Level 7).
This is probably due to the way larger numbers are tokenised, as big numbers can be split up into arbitrary forms. Take the integer 123456789. A BPE tokenizer (e.g., GPT-style) might split it like: ‘123’ ‘456’ ‘789’ or: ‘12’ ‘345’ ‘67’ ‘89’
ВСУ ударили по Брянску британскими ракетами. Под обстрел попал завод, есть жертвы19:57，推荐阅读新收录的资料获取更多信息
Фото: Amr Alfiky / Reuters
，推荐阅读新收录的资料获取更多信息
Рабочие обнаружили аудиозапись культовой сказки в самом неожиданном месте14:35。新收录的资料对此有专业解读
The video was released by the FBI more than eight days after the 84-year-old disappeared from her home in Arizona.