The trap most AI in recruitment falls into is treating every user action as equal training signal. That produces a model that gets louder every week, not sharper. Holi's learning loop is deliberately picky about what it learns from.
The signals Holi weights heavily
Rewrites. If you rewrite a draft substantially before sending, that is a strong signal about where the current model is wrong. Rewrites feed straight back into next week's drafts.
"Not this" rejections in the digest. When you tell Holi an observation is wrong, the underlying pattern gets marked down immediately. Three rejections on the same pattern kill it.
Rejected outreach recipients. If Holi suggested a person to reach out to and you said no, the reasoning behind that suggestion gets weighted down for similar profiles.
The signals Holi weights lightly
Opens. A good subject line does not mean a good message. Opens are useful in aggregate for subject-line patterns, worthless in isolation.
Accepts. A draft accepted unchanged might mean it was perfect. It might mean you were rushed, or the recipient was low-stakes, or you did not care enough to edit. Accepts are ambiguous; Holi treats them as such.
Time on page. Session-length metrics tell you nothing about whether the work was any good. Holi ignores them.
Why picky feedback matters
Because a model that treats everything as signal ends up drifting in whichever direction the noise points. Specialist recruiters need a model that changes when there is a real reason to change, and holds still otherwise. The learning loop is where the picky filtering happens, and the weekly digest is where you see what it decided.
The best feedback you can give Holi is a rewritten draft with a one-line note about why you rewrote it. That is worth more than a thousand opens. And Holi will show you exactly how that feedback moved the following week's defaults, section by section, in Monday's digest.
Sharper, not louder. That is what a working feedback loop looks like.