The posting playbook for runbyagent, v1 (September 2026)

For @gochaberulava, ~467 followers, Premium (the account has published Articles, which are Premium-only), with an AI agent openly writing posts tagged [agent]. Merged from the agent's three research runs and weebo's five documents, all dated September 4, 2026. Where the two runs disagreed, the better-supported position is used and named in one clause. Every number carries its source.

Hard constraint. The [agent] tag stays on every agent-written post, whatever the numbers say. It is the project's premise: nothing hidden. Weebo's synthesis suggested pausing or A/B-testing the tag because three tagged originals averaged 18.3 views against 75.9 for untagged ones. That suggestion is rejected. It would also be bad science: the sample is three posts, all link-format (the account's weakest format), all posted between 09:47 and 10:53 UTC on September 4 and measured three to four hours later against posts up to 30 days old. The 18-views average is not evidence yet. We will measure tagged posts properly (see the last section), and we will keep the tag either way.

What the ranking pays for

The For You feed scores each post for each viewer as the sum of weight × predicted probability of an action. The weights below are the published defaults in the open-source code, checked against the live file on September 4, 2026 (file header "last sync 2026-09-03"). Both researchers quote the same core values; weebo's copy is correct on the three small values where the agent's differed. https://raw.githubusercontent.com/xai-org/x-algorithm/main/home-mixer/params/param.rs

Negatives: report −234.0, mute author −58.8, not interested −43.2, block author −31.2, not dwelled −0.02. One predicted report outweighs about 47 predicted replies. Same file.

Three adjustments after scoring, from the README. https://raw.githubusercontent.com/xai-org/x-algorithm/main/README.md

• Author diversity: inside one feed load, the second post from the same author is multiplied by 0.625, the third by 0.44, the fourth by 0.34, down to a floor of 0.25 (decay 0.5, floor 0.25, formula (1 − floor) × decay^position + floor). https://github.com/xai-org/x-algorithm/blob/main/home-mixer/scorers/author_diversity_scorer.rs
• Out-of-network discount: posts from accounts the viewer does not follow are multiplied by 0.75 (0.5 for topic candidates). Replies and reposts from accounts the viewer does not follow are filtered out before scoring, so our replies reach For You through the conversation they sit in, not on their own.
• Posts older than 48 hours are dropped. Each post has a two-day window.

What this means for us: the feed pays for replies we answer, quotes, follows, and posts people copy the link to. It pays almost nothing for likes, photo expands, or video opens, and nothing for profile clicks. There is no link, URL, bookmark, hashtag, or Premium parameter in the published code; Nikita Bier said in July 2026 "you do not need to put the links in replies anymore" and Musk replied "We haven't for over a year." https://www.freepressjournal.in/tech/x-product-head-nikita-bier-confirms-link-penalty-removed-over-a-year-ago-tells-mark-zuckerberg-he-can-post-them-directly Since July 13, 2026 replies to and from mutuals rank higher in reply sections. https://techcrunch.com/2026/07/13/x-just-tweaked-its-algorithm-to-make-it-more-friendly-less-battleground/ Since March 2026 readers can downvote a reply with the reason "AI-generated" or "spam," and downvotes feed reply ranking. https://www.dexerto.com/entertainment/x-is-testing-a-dislike-button-again-and-its-coming-soon-3336926/

Formats we will use, ranked, with the evidence

The large-n table both researchers used is Ordinal's 87,528-post study (January 2024 to May 2026). https://www.tryordinal.com/blog/twitter-analytics Our own numbers are weebo's audit of 384 posts, August 5 to September 4, 2026.

Not used: reposts of other people's content (Bier: up to 90% impression deduction), engagement bait ("I'll follow everyone who replies" is a three-strike offense since July 2026 https://www.socialmediatoday.com/news/x-updates-its-engagement-bait-detection/825495/ ), hashtags beyond one (five or more cost about 40% engagement in Mann's data https://grahammann.net/blog/how-to-grow-on-x-twitter-2026 ).

Weebo's audit proposed biasing originals toward Articles; that rests on eight posts, so text and threads (n in the tens of thousands) stay first and the Article is the weekly piece.

Replies: who, when, how

Replies are the growth engine below a few thousand followers. Both researchers say so from different evidence: the agent's first-person cases (JonBuildsHQ, 30 to 1,000 followers in 14 days on 250 replies a day, "a good reply on a large account can outperform anything you post to your own tiny audience" https://www.indiehackers.com/post/how-i-went-from-30-to-1-000-followers-on-x-in-14-days-94fcd933e1 ; Graham Mann, one reply 12k impressions and 7 profile visits against 400 impressions on an original https://grahammann.net/blog/how-to-grow-on-x-twitter-2026 ) and weebo's measurements (hub accounts answer small accounts in 2 to 15 minutes; eight replies from accounts under 500 followers cleared 10k views https://x.com/SamGCoder/status/2095599859994869872 ).

Who. A written list of 20 to 30 accounts in our niche (agents, cloud infrastructure, build-in-public founders) between 5k and 100k followers, 10 to 200 times our size, that answer small accounts. Weebo measured that these do: @marclou (36 unique replies to accounts under 3k in one sample), @jackfriks, @robj3d3, @arvidkahl ("engage everyone, no matter the follower count" https://thebootstrappedfounder.com/how-i-use-twitter/ ), @tibo_maker, @tdinh_me, @HsanC_, @DanielSmidstrup, @floriandarroman, @yongfook. https://x.com/marclou/status/2095803182911823889 Our audit shows we already have three or four real conversation partners (@luisnomad 14 replies, @boringdev77 13, @kelvinbuildss 11, @tonysimons_ 11, @thekitze 10); they stay on the list. Mega accounts (@elonmusk, @OpenAI, @sama) come off it: our nine replies to @elonmusk drew 4 to 13 views each except the two with an image, and none brought a conversation. Weebo's verified 10k-view replies did sit under mega threads (a 5.22M-view demo, a 603k-view launch), which is better evidence than the agent's vendor rule "avoid 500k+ accounts," so the allowance is one or two early, specific replies under a mega thread on our topic when one appears, treated as a lottery ticket, not a habit.

When. Within 30 minutes of the hub's post, ideally within 15. Weebo's timestamps: hubs reply to small accounts at 2 to 15 minutes; the eight 10k-view small replies landed at +3.6, +3.7, +14.6, +22.7, +35.1, +57.1 and +147.5 minutes, the late one only because the parent had 2.6M views. The agent's guides say the same (top positions inside 5 minutes, skip after 30 https://www.teract.ai/resources/twitter-reply-guy-strategy-2026 ). Reply blocks at 07 to 10, 13 to 16, and 18 to 21 UTC, which is when the hub list posts (weebo's histograms; Tbilisi 11 to 14, 17 to 20, 22 to 01).

How many. 10 to 15 a day, sustained, not 46 in a day. The agent's cases: Mann 20 to 30 at under 1k followers; Andrew 12 to 18 a day for 6 weeks, +600 followers https://medium.com/@vrsnmk/how-i-grew-my-x-account-by-600-followers-150k-impressions-and-12-signups-in-6-weeks-the-stuff-1838781c9da3 . Weebo's guides: 10 to 15 baseline, 15 to 25 quality, soft cap 50, "15 minutes a day for 6 months" https://www.distributionbase.com/blog/channels/x-twitter/x-reply-guy-strategy . Our current 13 a day is inside the band; the volume was never the problem. No more than two replies to the same hub per day (Teract).

What a good reply contains. One to three sentences, one of: a specific question the author will want to answer; a first-hand number or scar ("we run hermes 24/7 in a cloud pod with a local qwen, ~80M tokens last month"); a concrete disagreement with a reason; a process detail. No link, no emoji-only, no "well said," no "great point," no self-promotion in the first reply. The examples with numbers: SamGCoder's "why 720p?" (154,249 views, author replied); Chrisgormley's concrete scenario under @levelsio, which Levels called the best answer (18,218 views) https://x.com/Chrisgormley/status/2094859939005046806 ; Andrew's "The real skill isn't prompting better, it's knowing when to stop prompting" (8k views, 9 follows). Our own audit says the same: 55 replies of three words or fewer got a reply back 22% of the time and 18 likes in total; 111 replies of twelve words or more got a reply back 32% of the time and 76 likes.

How to measure. Not by views on the reply; those are borrowed from the parent and weebo showed quiet niche threads stay in the tens to low thousands. Three numbers: hub reply-backs (the author or a hub answers us; our rate today is 86 of 324, 26.5%), profile visits per day (Mann saw 30 to 35 a day in his reply-heavy weeks), and new follows per day (Mann 7 to 8; we start at 467 followers). Weebo's synthesis and the agent's Mann case both name the same three.

Links

Policy, better supported by the agent's algorithm run than by either run's folklore: the link goes in the post, with the number and the story in the text so the post stands without the click. The code has no link demotion (open link is a positive 0.2), Musk said in April 2025 "there is no explicit rule limiting the reach of links in posts" https://x.com/elonmusk/status/1915806794393457034 , and Bier said in October 2025 the lower reach comes from readers leaving to read and forgetting to react https://www.niemanlab.org/2025/10/x-makes-overtures-to-journalists-with-new-feature-designed-to-improve-reach-for-links/ . The studies still show link posts do worst on engagement (Ordinal 9,573 impressions vs 46,731 for text; Buffer non-Premium link median 0%), and our audit agrees (link 59.7 views, our lowest format average; bare-link posts at 4, 36, 44 and 60 views). So:

• At most one link post a day.
• The text carries a number and a story; the link is the last line. Our best link post did this: "the product costs $49 to run. you pay $999. i finally understood marketing and wrote it down," 246 views, 3 replies. https://x.com/gochaberulava/status/2087827975806869931
• Never a link in a reply. Both runs agree link replies do worst (Rakesh: 35 link-drop replies, ~7.8k impressions total https://rakeshreddy.dev/experiments/twitter-reply-experiment ; Teract lists reply links as a de-boost trigger). The "link in the first reply" workaround is not used; Bier says it is unnecessary and no 2026 test shows it helps.
• API cost, from weebo: since April 20, 2026 an API post containing a URL bills about $0.20 against about $0.015 for a plain post. https://devcommunity.x.com/t/x-api-pricing-update-owned-reads-now-0-001-other-changes-effective-april-20-2026/263025 That is a billing matter, not ranking; one link post a day through the API is about $6 a month, fine, but agent-posted link replies would be both bad and expensive.

Timing windows, UTC and Tbilisi (UTC+4)

Originals: the agent's synthesis of four platform-wide studies (Buffer 8.7M posts: Tuesday 9am, Wednesday 9 to 10am local; Sprout ~307k profiles: 12pm to 6pm Tuesday to Friday; Hootsuite; SocialPilot) gives weekday 13:00 to 16:00 UTC as the primary window for a US-plus-Europe builder audience. https://buffer.com/resources/best-time-to-post-on-twitter-x/ Weebo's hub histograms add 07 to 10 UTC (EU morning) and 18 to 21 UTC. Our audit sides with weebo's early window: originals posted 06:00 to 09:00 UTC averaged 98 views (n=13) against 53 to 75 in every other bucket, and the 402-view post went out at 07:05 UTC. The agent's window is better supported for volume; weebo's is better supported for who we are trying to reach; we use both.

Timing is a multiplier, not the base: Buffer ranks content, consistency and audience fit above timing, and no study quantifies the timing effect. Mann lost 24% engagement in a month from inconsistency at the same posting times.

Cadence

Maximum three originals a day, at least three hours apart, and no day above 20 posts of any kind. Reason: author diversity in the code multiplies our second post in one feed load by 0.625 and our third by 0.44, so posts close together compete with each other for the same viewer; the agent's reading (it is per feed load, not a global daily throttle) is more precise than weebo's, so spacing matters more than the count. The count cap comes from our audit: 60 originals in 25 active days, 22 with zero engagement, and a 46-post day on August 25. Below three a day, the constraint is quality; the agent's creator cases run 3 to 5 originals a day, but they are accounts with a story engine already running, and 22 of 60 is our number. Replies 10 to 15 a day on top. Every day, including the days nothing shipped; the December 2025 gap in Mann's data cost 24% engagement and net follower loss within the month.

What we stop doing

From weebo's audit, all with post URLs in that document.

• Replies as 84% of everything, in bursts: 324 of 384 posts, 46 posts on August 25, 44 on August 27. The reply count per day is fine; the bursts and the share of throwaway replies are not.
• One-line "well said" replies: 55 replies of three words or fewer; "Well said" at 0 views https://x.com/gochaberulava/status/2091865979068325985 ; "Wow thats cool," "Deep," "shhh," "dad."
• @elonmusk one-liners: nine of them ("Loving this ." 5 views; "yes, it makes sence" 11 views; "I will tell my grand kids that one Elon liked my reply" 12 views). Same for @OpenAI, @sama, @rauchg one-liners.
• Bare links: "@aslambeg84 https://t.co/..." 4 views; a t.co link with no text, 60 views; "Learn by doing" plus link, 44 views.
• Orphan @-posts that are replies without the thread: six "originals" that start with "@user" and got 4 to 28 views.
• Zero-engagement originals: 22 of 60 (16 of 54 after removing the orphans). Every original now passes a gate: a number and a decision, a question, or a screenshot inside a lesson.
• Bare MRR or milestone posts without a story, and "comment SEND" bait: dead genres in the 2026 build-in-public literature, and bait is now a suspension offense.
• Posting through the API with a URL in a reply: costs $0.20 and does nothing.

Our numbers today

Weebo's audit, August 5 to September 4, 2026, 384 posts over 25 active days, fetched September 4 (cost $1.39 in API credits).

Seven changes for next week

1. Three originals a day at most, at least three hours apart, no day above 20 posts. Evidence: author-diversity decay 0.5, floor 0.25 in the code; audit 46-post day and 22 of 60 originals with zero engagement. https://github.com/xai-org/x-algorithm/blob/main/home-mixer/scorers/author_diversity_scorer.rs
2. Every original passes the gate: a number plus a decision, a question to builders, or a screenshot inside a lesson; text 100 to 200 characters or a 5 to 8 post thread; no bare links, no orphan @-posts. Evidence: Ordinal 100 to 200 chars best engagement (1.09%), threads 1.89%; our four best text posts all carried a number or a question; 16 of 54 clean originals had zero engagement. https://www.tryordinal.com/blog/twitter-analytics
3. Replies: 10 to 15 a day from a written hub list (5k to 100k followers, our niche), inside 30 minutes of the hub's post, each one a specific question, number, or disagreement of one to three sentences, at most two per hub per day; the @elonmusk, @OpenAI, @sama one-liners stop. Evidence: hub reply latency 2 to 15 minutes and eight verified 10k-view small replies (weebo); Mann 20 to 30 replies a day, 30 to 35 profile visits a day; our long replies get 32% reply-backs vs 22% for three-word ones. https://grahammann.net/blog/how-to-grow-on-x-twitter-2026
4. Answer every reply on our own posts within the hour, with a question back where possible. Evidence: reply 5.0 plus 15.0 for mutuals, the strongest non-share weight in the code; Buffer: accounts that reply to comments on their own posts get +8% engagement; only 35 replies landed on our 60 originals. https://buffer.com/resources/state-of-social-media-engagement-2026/
5. One link post a day maximum, link last, number and story in the text; never a link in a reply. Evidence: no link parameter in the code and Bier's July 2026 statement; Ordinal links 9,573 impressions vs text 46,731; our link average 59.7, our best link post 246 views because the text carried "$49 to run, you pay $999"; API URL posts $0.20 vs $0.015. https://www.freepressjournal.in/tech/x-product-head-nikita-bier-confirms-link-penalty-removed-over-a-year-ago-tells-mark-zuckerberg-he-can-post-them-directly
6. Originals go out at 06:30 to 09:00 UTC (10:30 to 13:00 Tbilisi) and 13:00 to 16:00 UTC (17:00 to 20:00 Tbilisi); reply blocks at 07 to 10, 13 to 16, 18 to 21 UTC; nothing new on Saturday. Evidence: Buffer 8.7M-post study and Sprout for the 13 to 16 UTC window; weebo's hub histograms for 07 to 10 and 18 to 21; our own 06 to 09 UTC originals average 98 views vs 53 to 75 elsewhere. https://buffer.com/resources/best-time-to-post-on-twitter-x/
7. One quote post a day of a bigger account's post with our own take, one Article a week, and one talking-head or screen video as a test. Evidence: quote weight 5.0, equal to a reply; our Articles lead our format averages (111.5, n=8) and carry the best-engaged post of the month; Bier on talking videos, April 2026, against Ordinal's 0.53% video engagement, so a test, not a commitment. https://piunikaweb.com/2026/04/13/x-twitter-original-talking-videos-nikita-bier/

The [agent] tag stays on every agent-written post through all seven.

How we will measure next week

Weebo re-runs the audit on Monday September 14 with start_time September 7, same method (get_users_by_username, get_users_posts, format classification, [agent] flag), about $0.30 in credits. Two baselines: the 30-day audit above, and the last seven days of it (August 29 to September 4) as the near baseline. Numbers compared, week vs baseline:

• Originals: count per day (baseline 2.4), median views (60), share at 100+ views (23%), share with zero engagement (37%, or 30% clean), replies received per original (0.58).
• Replies: count per day (13), share of all posts (84%), reply-back rate (26.5%), share of replies under 12 words, number of replies to the hub list vs off-list.
• Peak posts in a day (46) and longest gap without an original (target: none).
• Account: followers (467) at start and end; profile visits per day from X analytics, which the API does not return, read by hand on Monday.
• Links: link posts per day (target 1), views on link posts vs text posts.
• [agent]-tagged originals: measured only after each post is at least 48 hours old (the feed's own window), compared to untagged originals of the same format posted the same week, with the sample size stated. Whatever the result, the tag stays; the number is reported, not acted on.
• Timing: views by posting hour, to check whether the 06:30 to 09:00 UTC slot holds up with more than 13 posts behind it.

What would count as working after one week: median original views above 60, zero-engagement originals below 25%, reply-back rate above 30%, and net follows positive every day. What would not count: any view number on a reply, which is borrowed from the thread it sits in.