I asked Fable 5.1, Anthropic’s flagship model, to tailor my resume to a specific position. Nothing drastic: just make sure the skills that appear in the job description are specified verbatim without rewording, in case someone is doing keyword matching.

First, it didn’t get the point. Despite a clear order to include skills verbatim, it tried to get creative.

Second, it got thoroughly confused about some stuff. E.g. it thought that “subprocesses” are actually sub-agents, and that “callbacks” are Claude code hooks.

While moving things around, it misclassified them. For example, it categorized Pager Duty as an observability system.

The resulting resume looked utterly bizarre, I ended up fixing by hand.

On the other hand, Fabled helped me quite a bit with research and scoring of various job opportunities.

Also, Fable was very good at checking me once I finished my editing, down to missing hyphens and stuff like that. Does it mean I could have used adversarial review, like have another agent verify what the first agent did? The answer is no, since misclassification and confusion would not go away. There is no point spending tokens breaking the resume and then spending more tokens trying to restore it back to the original shape, with no guaranteed success.

So, research – yes, resume writing – not yet.