# Worth a Reply

> Tune a post or a thread for X against the engagement weights X published when it open-sourced its home-timeline
> heavy ranker (twitter/the-algorithm-ml, projects/home/recap/README.md, 2023). The page counts each post the way X
> counts characters, lists its links, mentions, hashtags, emoji and questions, flags the habits the source skill warns
> about, and charts the ten weights. A metered run says which engagements the draft invites or risks, gives a verdict,
> writes three rewrites of the first post - each aimed at a different engagement - and a plan for the replies. The page
> re-counts every rewrite and refuses one that adds a number, a mention, a link or a hashtag the draft never had.
> Nothing is posted from the page. Not affiliated with X. Built on the SkillSafe skill
> @composiohq/twitter-algorithm-optimizer.

https://worth-a-reply.skillsafe.ai/

## The published weights
score = sum of weight x predicted probability. Likes (fav) 0.5; reposts (retweet) 1.0; replies (reply) 13.5; profile visits that lead to a like or reply (good_profile_click) 12.0; half a video watched (video_playback50) 0.005; a reply the author engages with (reply_engaged_by_author) 75.0; a click into the conversation with a like or reply (good_click) 11.0; two minutes in the conversation (good_click_v2) 10.0; see-less, mute or block (negative_feedback_v2) -74.0; report (report) -369.0. The README says the weights were first set so each term adds about the same on average, and are adjusted over time: they show what the model counts, not what one engagement is worth today.

## What is free
- Posts: a line of --- or paragraphs numbered 1/ 2/ 3/ make a thread; anything else is one post.
- Counting, as twitter-text does: most Latin-script characters 1, CJK and other wide scripts 2, every emoji sequence 2 (skin tones and joined parts included), every link 23. The limit is 280.
- Findings: over the limit, very short, links, more than two hashtags, engagement bait ("RT if", "tag a friend", "follow for more"...), three or more words in capitals, three or more exclamation marks, more than five emoji, a post that opens with a mention, and a thin thread opener.
- Four samples (a launch post, a hot take, a question post, a four-post thread) replay saved tunes at no cost.

## The metered task
One task, `tune`, on gpt-terra, metered per run. Input string fields: `task`, `goal`, `audience`, `post` (the first post), `thread` (the rest, numbered, or none) and `facts`. The reply is one JSON object: `verdict` (`ready`, `rework` or `risky`), `signals` (each of the ten with `pull` - invites, neutral or risks - and `why`), `rewrites` (three, each with `aim`, `text` and `change`), `reply_plan` and `note`.

## Links
- API tutorial: https://worth-a-reply.skillsafe.ai/api.html
- Source skill: https://skillsafe.ai/skill/@composiohq/twitter-algorithm-optimizer
- The weights: https://github.com/twitter/the-algorithm-ml/blob/main/projects/home/recap/README.md