Threads Formula 30-Day Test, Wrap-Up: 13 AI-Written Posts, One Hit 100K Views, and the AI-vs-Handwritten Question We Can't Answer

Table of Contents

1. The Scorecard: All 13 Posts
2. The Verdict: Did Not Pass
3. Why the 100K Post Doesn't Count for the Formula
4. Reposts: Zero for 20 Days, Then 26 on One Post
5. The Opening Post's Third Question: AI-Generated vs. Handwritten
6. What Else We Didn't Get Right in These 30 Days
7. If We Ran It Again
8. How the Numbers Were Calculated

Part 7 of the "work system" series, and the wrap-up of the Threads formula test. The opening post (in Chinese) promised weekly data and a wrap-up once the window closed. Two things to own up to first: the week 3 update said the next one would come around August 31, and the week 4 update never shipped; the 30-day window closed on September 3, and this wrap-up is coming out in early November, about two months late. Both delays are ours, with no other cause.

Disclosure: Ultra Lab, MindThread and Ultra Advisor are products of the same group, Ultra Creation (傲創實業).

The conclusion first, in three sentences.

One: the formula did not clear the bar we set for ourselves. We published 13 stance posts in 30 days, and against the same accounts' everyday posts from the same period, 10 of the 13 landed below the median. That is not "consistently better."

Two: one post on day 28 reached 108,071 views, the only one of the 13 to pass 10,000. But the same account had already started taking off before it, and two weeks later a non-formula everyday post doing the same cigarette-cost math reached 118,988. That hit can't be credited to the formula, and the formula's contribution can't be ruled out either; a single post can't separate the two.

Three: the opening post asked how AI-generated formula posts compare with fully handwritten ones. We can't answer that. All 13 posts were AI-generated drafts, we did not schedule a single handwritten one, and the everyday posts we used as the baseline were AI-generated too. The handwritten leg was missing from the design on day one.

The details follow.

#The Scorecard: All 13 Posts

The window runs from the evening of 2026-08-04 (the first formula post on universe) to the end of 09-03, Taipei time. View counts come from a re-pull on October 31, so posts that appeared in the weekly updates now show slightly higher numbers (week 3's 312 and 284 are now 316 and 325). Views keep accruing after publication; these are not corrections.

Published (Taipei)	Account	Views	Reposts	vs. pre-launch median	vs. same-window everyday median
Aug 4	universe	494	0	+35%	+21%
Aug 5	risk	283	0	+125%	−18%
Aug 13	risk	64	0	−49%	−81%
Aug 13	risk	54	0	−57%	−84%
Aug 14	risk	66	0	−47%	−81%
Aug 14	universe	170	0	−54%	−58%
Aug 19	universe	316	0	−14%	−22%
Aug 23	universe	325	0	−11%	−20%
Aug 24	risk	2,558	1	20.4×	7.4×
Aug 27	universe	205	1	−44%	−50%
Aug 28	risk	232	0	+85%	−33%
Aug 31	universe	248	0	−32%	−39%
Sep 1	risk	108,071	26	861×	314×

The accounts are @risk.clock.tw (risk) and @universe_signal_tw (universe). Percentages show how far above or below the median a post landed; "×" is views divided by the median. The two yardsticks:

• Pre-launch median: every post from the same account between July 5 and launch on August 4. risk: 22 posts, median 125.5. universe: 121 posts, median 367.
• Same-window everyday median: every non-formula post from the same account inside the 30-day window. risk: 52 posts, median 344.5. universe: 116 posts, median 407.

Baseline medians also shift slightly with each re-pull; risk's pre-launch median, reported as 121 in weeks 2 and 3, is 125.5 this time. One more thing to disclose: both of universe's yardsticks include video posts (23 pre-launch, 24 in the window, with medians of only 117 and 118), while all 13 formula posts are text. Comparing text against text only, universe's pre-launch median becomes 407 (98 posts) and its same-window median 490.5 (92 posts); the best of its 6 formula posts, the first one, drops to +21% and +1%, and the other 5 all fall below the median. This doesn't change the conclusion; it only makes universe's results look worse. risk's everyday posts in the window are all text, so it is unaffected.

One more universe formula post went out on the evening of September 4 (137 views). It falls outside the window and is not in the table.

#The Verdict: Did Not Pass

The opening post set the bar: formula posts need to beat the same account's everyday posts by 50% or more, consistently, before the formula is worth teaching (from week 1 on, measured against the median).

On the pre-launch yardstick, 4 of 13 cleared +50%, which looks decent. But all 4 are risk posts, and risk's pre-launch baseline is only 22 posts from a stretch when the account posted little and drew few views. Once it went back to daily posting in August, the whole account's level rose. We flagged this in week 1; using it to declare a pass would be unfair.

On the same-window everyday yardstick, only 3 of 13 beat the median, 2 cleared +50%, and the other 10 fell below the median. None of universe's 6 posts cleared the bar; its best was the first one (+35% pre-launch, +21% same-window).

Even with the three bug-affected posts from Aug 13 to 14 set aside (the ones week 2 said it would track separately), 7 of the remaining 10 still fall below the same-window median.

The two yardsticks need keeping apart. The pre-launch one is the baseline fixed in week 1, and 4 of 13 posts cleared +50% on it. The same-window yardstick was added later, and 2 posts cleared it. Whichever you use, at most 4 of 13 is not "consistent." By the bar we set in advance, the formula did not pass on these two accounts.

The opening post's second question, whether the effect decays with repetition, gets a plain answer too: we can't tell. universe's 6 posts ran 494, 170, 316, 325, 205, 248 in order. No downward trend, no upward one, and 6 posts can't carry a trend claim either way.

Week 3 promised to check the scheduler fix in practice, so here is that result: after the fix shipped on August 17, both accounts published formula posts exactly four days apart (risk on Aug 24, 28 and Sep 1; universe on Aug 19, 23, 27 and 31). The back-to-back bursts of August 13 did not recur.

#Why the 100K Post Doesn't Count for the Formula

The September 1 risk post was a cigarette-cost calculation: a pack a day at NT$150 is NT$54,750 a year and NT$547,500 over ten years. As of October 31 it had 108,071 views, 1,112 likes, 311 replies and 26 reposts. That is 314× the same-window everyday median, and the highest of all 59 risk posts in the window.

Looked at alone, this post invites a "the formula finally caught fire" headline. Laid out against the whole account, the picture changes:

• risk looked like a different account after August 25. Its 34 everyday posts from Aug 4 to Aug 24 had a median of 278 and a high of 1,095. The 14 everyday posts from Aug 25 up to the Sep 1 formula post had a median of 632, including 11,805 on Aug 25 and 26,217 on Aug 27, both before the formula post. From the September 1 post's point of view, the account took off first and the formula post came after. In fairness, though, the first risk post to break the old ceiling was the Aug 24 formula post (2,558), and everyday posts started passing 10,000 the next day. Which one started the takeoff can't be separated.
• The cigarette-cost topic is no guarantee on its own. Before the account went quiet in July, risk's cigarette-cost everyday posts drew between 156 and 1,051 views. On August 26, after the takeoff and before the formula post, a non-formula version drew 6,881.
• After the window, on September 14, risk published a non-formula everyday post doing the same math (NT$54,750 a year, about NT$1.1 million over twenty years). It reached 118,988 views and 38 reposts, more than the formula post. Through October 31, risk published 5 cigarette-cost everyday posts after the window, and 4 of them passed 10,000 views (Sep 14: 118,988; Sep 28: 81,531; Sep 16: 55,110; Oct 22: 23,795). Those later posts are nearly identical in content; we haven't checked whether their generation was influenced by the Sep 1 post.
• Cigarettes weren't the only topic taking off. Of the 7 everyday posts risk published after the window that passed 50,000 views, 3 were about cigarettes and the other 4 were long-run cost calculations for coffee, a scooter, cocktails and a commuter car. The first two everyday posts in the window to pass 10,000 (Aug 25 and Aug 27) calculated the ten-year cost of owning a car.
• In the same stretch, formula posts also lost: the Aug 28 post on credit card interest drew only 232 views during the account's hottest week, −33% against the same-window median.

After restarting on August 5, every one of risk's 52 everyday posts in the window cites a specific money amount, and all 7 of its formula posts use the same convert-spending-into-money approach; by format alone you can't tell the formula posts apart. Our reading is that the post took off mainly because the account was taking off at the time, with cigarettes as one of the topics that happened to land, and the formula's structure helping at most along the way. It is only a reading: the window has exactly one formula-written cigarette post, and the later non-formula ones went out at different times with the account in a different state, so comparing them proves nothing.

One smaller observation: the "stance" in this post is thin. The formula calls for naming an opponent and saying what you did instead; this post is mostly a string of calculations and looks a lot like risk's everyday posts. Its success doesn't look like the stance structure's doing.

#Reposts: Zero for 20 Days, Then 26 on One Post

The formula's bet is that a stance triggers reposts and reposts drive reach. The record:

• The first 8 posts (Aug 4 to Aug 23) got zero reposts, all of them.
• The post published on day 20 (Aug 24) later got 1 repost, and the Aug 27 universe post got 1.
• The Sep 1 post alone got 26.
• Across all 13 posts: 28 reposts, 26 of them on a single post; 10 posts had none.

For comparison: of universe's 116 everyday posts in the window (videos included), 29 had at least one repost, 45 reposts in total; counting only its 92 text posts, 22 were reposted, 36 reposts in total. Its 6 formula posts had 1. The fuel was there all along; on this account, the stance posts never lit it. risk went the other way: its 52 everyday posts in the window had 2 reposts, its 7 formula posts had 27. But 26 of those came from the single Sep 1 post, which happens to end with the formula's "invite a share" step, and the non-formula cigarette post on Sep 14 drew 38. One post can't carry a conclusion, and the two accounts point in opposite directions.

#The Opening Post's Third Question: AI-Generated vs. Handwritten

Straight answer: we can't answer it, because there was no control group.

Where exactly it falls short:

1. All 13 formula posts were AI-generated through MindThread's production generation flow. We did not schedule a single handwritten formula post in the 30 days.
2. The everyday posts we compared against weren't human-written either. All 168 everyday posts on the two accounts in the window are marked in the database's source field as MindThread scheduled auto-generation. What the 30 days actually compared was "AI writing to the formula" versus "AI writing in the account's usual style." Handwriting never entered it.
3. This wasn't cut halfway through. The four design points in the opening post (formula built into the product, fatigue cooldown, real accounts, metrics fixed in advance) never included a handwritten leg. The question made it into the opening post; the design never scheduled it. We missed it at design time.

The opening post also said a human would only do the final review. What that meant in practice should be stated. The last batch (generated Aug 24) put 5 posts into the window, and the published text matches the model output word for word. The human part was selection: that batch ran four generation rounds, all 12 drafts from the first two rounds were discarded, and the posts used were picked from the last two. The first 2 posts were similar: we weren't happy with the first version, the whole batch was regenerated, and what went out was the text finalized in July. For the 6 posts from the Aug 12 batch in between, no draft files were kept, so we can no longer check whether any words were changed. "Review" in this experiment meant choosing drafts more than editing them.

Without a control group, only two things can be said:

• AI-written posts aren't kept from taking off just because AI wrote them. The 100K cigarette post went out exactly as the model wrote it.
• The length rule wasn't held at first. The formula sets a length of 60 to 150 characters. Not counting spaces and line breaks, 5 of the 13 posts ran over, all of them the first 5 (Aug 4 to midday Aug 14), the longest at 184; the remaining 8 all stayed within 150. Counting spaces and line breaks as well (the method behind the 179 figure in week 1), the first 5 still all ran over, and 4 of the later 8 ran slightly over (154 to 164). We adjusted the generation instructions and method along the way and kept hand-picking drafts throughout, so which change did the work can't be separated.

What can't be said: whether a handwritten version would do better or worse, or by how much. We don't have that number and won't guess.

#What Else We Didn't Get Right in These 30 Days

All listed here, because they belong next to the numbers:

• The formula itself changed midway. The original layout rule was "one block, no line breaks." We dropped it on August 12 and switched to requiring paragraph breaks above 100 characters. Strictly speaking, the first 2 posts and the last 11 used different versions of the formula.
• On Aug 13 to 14, a bug in one of the scheduler's fallback rules made risk publish 3 formula posts within 25 hours, defeating the fatigue cooldown. Those were the 3 worst posts of the run (54 to 66 views). The fix shipped on August 17.
• risk's second restock slipped, giving week 3 zero samples. The week 3 update already owned up to it.
• The week 4 update never shipped and this wrap-up is about two months late, as stated at the top.
• The opening post said we'd track net follower change per post. We didn't track it per post and didn't snapshot follower counts on September 3, the day the window closed. All we have: risk went from 1,513 on Aug 10 to 1,578 on Oct 31, universe from 280 to 428. That span includes two months of posts after the window and can't be credited to the formula.
• The opening post said the stance had to be real. All 13 stances were written by AI for the accounts' personas.
• View counts are an October 31 snapshot, nearly two months after the window closed. Earlier posts have had longer to accrue, which affects formula and everyday posts alike, and we have no daily data to correct for it.

#If We Ran It Again

The biggest takeaway is knowing what a test worth trusting actually costs. If we ran it again:

• The handwritten leg goes in on day one. Same account, same topics, half written by AI to the formula and half written by hand, posted alternately. Otherwise differences in topic and account state swallow everything, and the cigarette posts are the example here.
• Every generated draft gets archived, and every human edit is visible, so "review" can be measured.
• The account has to be stable first. risk went quiet in mid-July, restarted on August 5 and took off from August 25; an account like that can't measure a formula with any precision.

#How the Numbers Were Calculated

• Formula post list: every document in MindThread's content library tagged with the stance formula. 14 were published, 13 inside the window.
• Views, reposts, likes, replies: per-post insights from the official Threads API, pulled on 2026-10-31.
• Medians and percentiles, not averages (we changed this publicly in week 1 and kept it since).
• Day counts start with Aug 4 as day 0.
• The pull and calculation scripts live in our internal version control. They only read from the APIs and write nothing.

The formula templates and fatigue cooldown are still in the MindThread formula marketplace, but this test did not show them beating the accounts' usual content. The opening post promised: if it replicates we say so, and if it doesn't we say that. The answer this time: on these two accounts, we did not see it consistently beat everyday posts. And the question we most wanted to answer couldn't be answered by this design. Next time starts with the control group.

---

Earlier in the series: opening post (in Chinese), week 1, week 2, week 3. The original methodology is open-sourced under the MIT license; our templates are a rewritten implementation.