Projects / Case study

Reply Desk

Finds X posts worth replying to and drafts a reply in my voice, for about a dollar a day. It never posts: every reply goes out through my own tap.

Role
Solo builder
Timeline
7 to 12 November 2026
Built with
Python, NiceGUI, Postgres, Claude Haiku and Opus
Status
Private, runs on my home box

The problem

The way to be seen on X is to reply well to people in your field. Doing that means scrolling for the right posts, and most of what you scroll past is spam, bots or posts with nothing to reply to.

I wanted the good posts found for me and a first draft waiting, without handing my account to a bot. X suspends accounts for automated posting, and I didn't want my name on replies I hadn't read.

What I built

Reply Desk checks the accounts and searches I choose, scores each new post, drafts a reply to the good ones and puts them on a dashboard I open on my phone or laptop. It never posts. It holds a read-only X token, so it can't post even by mistake. Every reply goes out through my own tap in X.

Free rules go first, in code. They hide posts from accounts without a checkmark, posts that are only a link or under five words, text another account posted the day before, and posts with too many hashtags, capitals or fancy letters. Claude Haiku scores what's left from 1 to 10 with a short reason, and only posts that score 6 or more get a draft.

What one check does with postsExample posts, made up
  • We cut our eval bill in half by caching the judge’s answers. Curious whether anyone else scores with a smaller model first?Score 8: asks a real question about evals, in my field
    Drafted
  • Shipped a new onboarding flow today. Retention numbers next week.Score 4: nothing to add to it
    Hidden
  • FREE AI COURSE!!! LINK IN BIO #ai #ml #freeRule: mostly capitals, too many hashtags. No call to Claude.
    Rule
  • So true.Rule: under five words
    Rule
  • RuleHidden by a free check in code, before Claude sees it.
  • HiddenClaude Haiku scored it under 6.
  • DraftedScored 6 or more, so it gets a draft and a card.

Drafts I'd actually send

A draft picks up one thing the author said and asks them one question about their own work, then ends with a reason to connect. The code checks every draft: under 280 characters, no dashes, no emojis, in English, and none of a list of phrases that claim an experience I haven't had, like “I’ve seen” or “we tried”. A draft that breaks a rule goes back to Claude with the reason, up to three tries.

Then it learns from me. Each time I edit a draft and send it, the edit is saved. Every draft is written with my ten newest edits as examples, and after every five edits Claude Opus rereads all of them and rewrites a short list of lessons, such as what I always cut.

Keeping it on a budget

X charges for every post the app reads, so the budget lives in the code, not in a dashboard I might forget to check. The poller never asks for posts more than 12 hours old, caps each request at what's left of 200 reads a day, and stops until midnight UTC once they're spent. Between 07:00 and 23:00 it checks by itself if nobody has checked for two hours. It never checks overnight and refuses a manual check within 10 minutes of the last one.

What each step costs
StepRuns onCost
Read a post from XX APIabout $0.005 a post
Spam and checkmark rulesCodefree
Score a postClaude Haiku 5.5about $0.0002
Draft and check a replyClaude Haiku 5.5about $0.0004
Rewrite the lessons, every 5 editsClaude Opus 5.5about $0.10
Daily cap: 200 readsabout $1 a day
Costs from the code and the X price list. Reading from X costs about ten times more than anything Claude does, which is why the cheap rules run first and searches only ask X for verified authors.

How I test it

A hundred tests cover the poller, the scorer, the drafter and the dashboard. The database tests run against a real Postgres in CI. For the drafting prompt I built try-prompts: it saves 15 posts, drafts a reply to each with every prompt version I give it, and writes a page with the replies side by side. A comparison costs about $0.003 a version, so I compare before I change the prompt, not after.

What's next