Design decision 0007 · accepted
A free first look on the Mac before spending tokens
Apple's on-device model gives a quick score in seconds and skips obvious bait, calibrated against Claude on real posts.
Context
Some saved posts are bait (“comment AI for the link”) or off-topic. Sending every one to Claude wastes tokens.
Decision
When the post was read in full, the on-device Foundation model gives narrow sub-judgments (specificity, relevance, novelty, self-promotion, bait). Code combines those with text signals into a score:
- junk scores 2
- everything else gets a conservative 4–7
- only Claude can give 8 or more
Junk skips Claude unless you tap “Stash anyway”. App-building knowledge (pricing, onboarding, App Store, growth) counts as relevant even without agent vocabulary.
Consequences
- Calibrated on 22 real posts: average error 0.9 against Claude (1.2 on held-out posts).
- When the on-device model is unavailable, Stash simply goes straight to Claude.