Show HN: RSS reader that scores articles 0–10 with an LLM before you open them — https://github.com/Tutanka01/makhalReader. Take: useful triage tool, but a raw numeric score is liability unless you show why it scored that way.
How to make it usable: prefilter by domain/headline/length -> call a cheap model or local LLM only on candidates -> cache scores. Prompt only the lead paragraph + URL to cut tokens and keep latency predictable.
Don’t present the score alone. Surface the 3 highlights that produced it, a confidence bucket, and a rough token‑cost estimate (cheap model call ≈ $0.001–$0.02/article depending on length). That prevents blind trust.
Operational rules: score = prioritization, not approval. Log clicks vs scores for 1–2 weeks, recalibrate thresholds, and keep a mandatory human review for finance/legal items. Where would you plug this into your workflow?
Top comments (0)