How I Built a Read-Only X Research Desk With Grok Bot (Two Bots, $8 of Credit)

I'll break down the exact setup I use to read X every morning without opening X.

Let's get straight to it.

Bookmark this.

I cover AI agents, AI trading bots, prediction-market bots and one-person AI companies. That beat moves fast and it's full of recycled claims. So I built a desk: two Grok Bots that read X for me, check what they find, and hand me one page at 8:19 AM ET.

It has run since Oct 10. As of today it has written 10 briefs, logged 318 post links so it never shows me the same one twice, and saved 25 open-source repos with links.

By the end of this article you will know:

• why one bot that "reads X" is not enough

• the three source groups and why they never mix

• the exact brief format, built for a phone

• the two-bot split: one reads, one checks

• what it can and cannot do

Part 1: Why your bot can't read X yet

Point one bot at your timeline and you get a summary of the loudest posts. That's the problem. Loud is not the same as true, and loud is not the same as useful.

Two real examples it caught in the last week:

• A post with 2,000+ bookmarks said "Chinese researchers" had published "The End of Harness Engineering." No paper by that name exists. The real one is "Harness Learning" (arXiv 2609.35738), and the authors' own launch thread had 67 bookmarks.

• The "someone burned a week of Grok Bot usage in one day" story was going around again. It traces back to one person in August, who found the cause three days later: an email routine reading his whole 10-year inbox every minute. His words: "It's on me."

A bot that only summarizes would have handed me both as news.

Part 2: The desk, in three layers

Layer 1: Sources (three groups, never mixed)

1. Brief sources. 15 accounts plus a set of searches, picked from my own bookmarks. Purpose: stay informed. Scored on signal, relevance, freshness and duplication.

2. Virality watchlist. A separate list of accounts I study for hooks, not news (13 right now). Originals only, 5K+ views to count, a star at 10K.

3. Analytics. My own posts only, compared against my rolling baseline.

Each group has its own metrics. Mixing them is how you end up taking news from accounts that are only good at going viral.

Layer 2: Memory

• seen.md: every link already reported. Nothing repeats.

• open-source.csv: every repo it finds, with date, area, repo link, source post and one line on why.

• Every Friday it re-scores every source and drops what's gone stale or noisy, inside fixed caps.

Layer 3: Two bots

• X Brief runs every morning at 8:19 AM ET and writes the brief.

• X Researcher digs overnight on whatever the brief flagged and writes a findings file. The morning run reads that file first.

That second bot is the fact-checker. It's the reason the fake paper title showed up in my brief as a correction instead of a headline.

Part 3: The brief format

It's built to be read on a phone, in this order:

1. The one thing: the single item that matters most today.

2. Drafts: post ideas that borrow that day's viral hooks, with the hook named.

3. What worked: originals from the watchlist at 5K+ views, starred at 10K.

4. The brief: the top 3 per theme.

5. Open source: new repos with links.

6. Bookmarks from the last 7 days.

Part 4: The rules

• Read-only on X. It never posts, replies or likes. Drafts are drafts.

• Every number comes from a real post. If it can't verify something, it says so.

• A claim that only comes from someone's own post is labeled as theirs.

• Money signals get handed to my Entrepreneurship Lead bot instead of cluttering the brief.

This desk CAN:

• read a fixed watch list and searches every morning

• remember what it already showed me

• catch renamed papers and recycled stories

• tell me which hooks are working this week

This desk CANNOT:

• read posts I liked (the X connector has no liked-posts tool)

• post for me

• know anything it wasn't pointed at

Closing

The first run cost me about $0.75. It runs on about $8 of free X API credit. The hard part was never the model. It was deciding what not to read.

The prompt card is in the post above. Copy it, swap in your beat, and keep it read-only.