Fact-Anchored Replies
Hunt reply drafts now default to silence. The generator writes only when your bound brain holds a concrete fact that bears on the tweet, surfaces that fact next to the draft, and skips everything else instead of manufacturing filler.
Most reply tools always produce a reply. Hand them a tweet and an intent and they return a draft, every time, even when there is nothing real to say. With no substance to draw on, the only shape left is the slop template: validate the tweet, add a generic observation, pivot to the product. That brands you as the reply guy everyone scrolls past.
Fact-Anchored Replies fixes the cause, not the phrasing. The generator now defaults to silence and speaks only when it is holding a concrete fact.
How it works
When a tweet matches one of your Hunts, the generator runs one extra step before it drafts anything. It queries the brain bound to that Hunt for a single specific fact that genuinely bears on this tweet: a number, a benchmark, a war story, a correction, a thing that bit you. That fact is the anchor.
- No anchor, no reply. If the brain holds nothing concrete for this specific tweet, the generator returns
skip — nothing to addand writes no draft. Skipping is a real outcome, not a failure. Vague topical overlap does not count as a fact. - Anchor found, anchor leads. When a fact qualifies, the draft is built around it. The reply leads with the useful thing and validates nothing. No "great point," no "exactly this," no restating the tweet back.
- The fact rides with the draft. Every produced draft shows the anchor fact and the note it came from, so you can see exactly what the reply is standing on before you ship it.
Because the gate is strict, a mixed batch of targets produces real skips. Changelog bots, off-topic acronym matches, and tweets your brain has nothing to say about get withheld instead of answered with filler. The drafts that survive are the ones worth sending.
Pitching is rare and opt-in
The anchor fact is always the spine of a reply. A product mention is never the reason a reply exists.
Product and link mentions are off by default. A draft can only include one when you explicitly pass the pitch flag, and even then they are rate-limited to roughly one in twenty drafts across a Hunt. A reply whose only justification is the pitch never goes out.
Brand voice, applied last
After the draft is written and refined, it gets one final pass against your canonical brand voice from get_brand_voice. This adjusts wording only. The anchor fact, the stance, and the opening all stay put. Tone and house style rules are the last thing applied before the reply reaches you.
Measuring the withhold rate
Each Hunt tracks how often it speaks and how often it stays quiet:
replies_drafted— drafts the generator produced.replies_skipped— targets it withheld for lack of a concrete fact.pitches_sent— drafts that actually carried a product or link.
A skip rate above zero is the point. It proves the generator is withholding, which is the whole reason the drafts that do land are worth your time.