The story of a salaryman who failed at Threads side hustles 3 times, and found the real cause of death after having an AI perform an 'autopsy'
I'll post my report card first
Accounts attempted: 3
Duration: 2 months
Final follower count: 12
Revenue: 0 yen
Expenses: 24,800 yen for study materials
Yes, it's a huge loss
Aiming for '50,000 yen a month with Threads x Affiliate Marketing', I launched 3 accounts in 2 months, and I watched all 3 of them die
This isn't a memorial service for them, but rather a story about how when I had an AI perform an 'autopsy' on all three at once, it revealed a cause of death completely different from what I had imagined
This is a story about how, when I had an AI perform an 'autopsy' on all three at once, it revealed a cause of death completely different from what I had imagined
For those who are also thinking, 'My design is perfect, so why isn't it growing?', this will probably hit home
Account 1: The account that died from design alone
The first account was a gadget review account with a cynical character, intended for X (formerly Twitter)
For this account, I wrote up an operational design plan, an LP structure proposal, post examples, a revenue roadmap, and even a 'complete automation manual'
And I wrote up an operational design plan, an LP structure proposal, post examples, a revenue roadmap, and even a 'complete automation manual'
……And this was while I had 0 followers
I was writing an automation manual on the very first day of design
It's like planning the introduction of an unmanned register before even opening the store
So, I ended up abandoning it after 2 weeks without really posting anything
The trigger was finding and buying 'better-looking study materials' (24,800 yen)
My diary said this:
'Archived due to failure. Factor analysis to be done at a later date'
This 'later date' came 48 days later
I had completely forgotten about the factor analysis until the AI pointed it out
Account 2: The account that died of rage in 3 days
The second account was a character on Threads that introduced casual time-saving goods
This time, I properly wrote it down in the design document.
"Absolutely no affiliate links for the first 3 days"
"Limit posts to 1-2 times per day"
And here is the actual log.
85 posts in 3 days
Moreover, there were 8 consecutive posts between 2:47 AM and 3:00 AM,
which showed machine behavior no matter how you look at it.
(I was using a tool to post in bulk. Humans don't post every minute at that hour.)
Naturally, I also included links. On the third day.
As a result, the account went silent.
Reach became almost zero, and it was out of the game in 7 days.
I broke the rules I wrote myself using my own tools and died.
I completely misused the benefits of civilization.
Account #3: The most promising, yet the most severe case.
Account #3 was my main hope.
I incorporated all the lessons from account #2,
thorough PR labeling, links in the comment section, no consecutive posting,
and a design document that even explicitly stated, '80% of activity should be used for replying to others.'
The self-introduction in the first post got 7 likes.
Oh, this might actually work.
......From there,17 days of no posts.
What I was doing was building an automation tool to manage posts.
It was a system where post drafts would arrive on my smartphone's LINE, and I could approve them with a single button.
(This part actually worked well; it was a good system, if I do say so myself.)
And after 17 days of silence, as if to make up for it,23 posts in 4 days,
with 35% of them containing affiliate links.
From the perspective of the Threads algorithm,
it was a textbook example of spam bot behavior: 'long-term neglect -> sudden consecutive posts -> full of links.'
Reach died, and even after I sealed all the links and continued posting, it didn't recover.
On the 34th day, I closed it with 12 followers.
AI autopsy: The cause of death was neither 'lack of effort' nor 'lack of know-how'.
On the day I laid account #3 to rest,
I gave the AI agent all the files from the past two months and asked it.
"I don't need encouragement or generalities. Just identify the cause of death."
The post log timestamps, design documents, diaries, analysis reports,
this is the autopsy result that came out after cross-referencing everything.
Cause of death: 80% was 'creation work', 20% was 'contact with the market'
I have piles of design documents, manuals, and tools left over
On the other hand, posts, replies to others, and recording numbers are sparse
The clincher was the spreadsheet I made for tracking
For two months, not a single line of data was entered
In other words, for two months, I thought I was doing 'SNS management' but I was actually 'developing an SNS management system'
I thought I was doing 'SNS management'
I was doing 'developing an SNS management system'
Writing blueprints is fun. Tools always get finished
But reactions to posts are uncertain, replying is tedious, and recording numbers is boring
So my hands kept drifting toward what was fun and certain
It didn't just fall on deaf ears; I thought my eardrums were going to burst
And there was one more fatal wound common to all three generations
The characters I created were introducing products they hadn't even used
There isn't a single photo of the actual product
The timeline is just a list of paraphrased catalog specs
No wonder nobody's fingers stopped scrolling
So, what now?
As a result of the autopsy, I decided not to create a fourth one immediately
Instead, the first thing I did was
write down '10 Rules of What Never to Do Again' and 'Numerical Exit Criteria' as a constitution
My motivation betrayed me three times, so I decided not to trust it anymore
I only trust the system
From here on, I will disclose the specific procedures of that autopsy and the entire content of the 'constitution' I created
I will disclose the entire content of the 'constitution' I created
What instructions I gave the AI to perform the autopsy (the structure of the prompt)
How to create a 'Cause of Death Matrix' that breaks down the causes of death for the three generations
10 Rules of What Never to Do Again (full text and the failures that became the basis for each clause)
10-item checklist before launching
A numerical trigger table that decides in advance 'when to withdraw'
This content is for 'people who have failed once and are wondering why' rather than 'people who are just starting'
1. How to perform an AI autopsy (prompt structure)
The point is to activate the AI not as a 'comforter' but as a 'coroner'
I'll leave the skeleton of the instructions I used here
役割: あなたは私のSNS運用3連敗を検死解剖する分析エージェントです
励ましや一般論は不要。死因の特定と再発防止策だけを出してください
分析対象: (設計書・投稿ログ・日記・分析レポートのファイル一覧)
手順(この順で。いきなり結論を書かない):
1. 時系列復元 — 各アカウントの開始→意思決定→撤退をファイルの日付から復元
2. 設計比較表 — ペルソナ/導線/投稿の型/自動化度を3代並べる
3. 死因判定 — 全代共通の死因と、その代固有の死因を分ける
4. 学習曲線の評価 — 「直った失敗」と「繰り返した失敗」を分ける
※繰り返した方が本当の病巣
5. 反証チェック — 「そもそもこのジャンル自体が無理筋では?」も検討させる
ルール: 全ての主張に根拠ファイルを添付。事実と推測を分けるThe key to this prompt is step 4
Anyone can reflect on a single failure
But 'lining up three failures and extracting only the parts that haven't been fixed' is
a task where human self-justification gets mixed in and accuracy drops, and
this is the part worth leaving to AI
Also, if you don't include the 'counter-evidence check' in step 5,
the AI will lean toward answers that assume you'll continue, saying
'Your method is correct, just fix the operation'
It's important to put the option of withdrawal on the table from the start
2. How to create a cause-of-death matrix
Place the four categories of 'Concept/Content/Algorithm/Operation' on the vertical axis,
place each account on the horizontal axis, and
always paste 'evidence logs' into the cells
The rule is to fill it with evidence, not impressions
In my case, this matrix was the first time
'rows with the same content written in all three columns' = the disease focus repeated three times
was visualized
3. The 10 Commandments of Never Doing It Again (Full Text)
I'll post them exactly as they are from my actual constitution
The ( ) after each clause is the failure that gave birth to that clause
Do not post any external links for 30 days after opening(35% links with zero buff -> shadowban)
Do not post more than twice an hour. Up to 3 times a day until day 30(85 posts in 3 days, continuous posting every minute late at night)
Do not post about things you haven't used or experienced. Do not handle subjects for which you cannot provide actual photos(2 months of catalog posts with zero actual photos)
Prohibit reusing hook templates. If the first 10 characters match a past post, it's rejected(Reusing template collections in rotation caused AI-like smell to be inherited)
Do not have weeks with zero measurement. Weekly reviews without numbers are invalid(Measurement sheet had 0 rows for 2 months)
Prohibit new tool development for 4 weeks from the start of operation(Spent 17 days after opening on tool development and went silent)
When breaking a rule, amend the rule in writing before breaking it(Deviating 'just for today' was the cause of death for all three attempts)
Make a decision within 48 hours once a withdrawal trigger is pulled. No life support based on mood(None of the three attempts had withdrawal criteria, leaving death to external events)
Complete the autopsy report for the current project before moving on to new materials or methods(I left the factor analysis for 'later' and abandoned it for 48 days)
Design operations to run in 15 minutes a day. If it doesn't work, scale down rather than automate(The 'dream of full automation' led to over-engineering and abandonment three times)
4. Pre-launch Checklist
Check in a list whether you can prepare 'actual experience, real photos, and real data' for the first 30 posts
Is the persona an extension of 'what you are actually doing'? (Acting won't last for months)
Take 10 real photos in advance
Create a 30-day no-link calendar
Save the withdrawal criteria with a date prior to the launch date
Rehearse recording to the tracking sheet once
Turn on notifications for 10 accounts to engage with
Create the first week's 7 posts and run them through an AI-smell check
Let the account sit for 1-2 weeks after creation before using it
Fix the day of the week for weekly reviews on your calendar
5. Withdrawal Criteria (Numerical Trigger Table)
These are the actual figures I will use next. Please adjust the numbers to fit your own situation.
What matters is not the accuracy of the numbers, but 'deciding them before launch and putting them in writing'.
| Point in time | Metric | Trigger | Action |
|---|---|---|---|
| 2nd week | Weekly impressions | Less than 1,000 | Replace all post templates |
| 4th week | Followers | Less than 50 | Redesign or decide to close within 48h |
| 8th week | Followers | Less than 200 | Close (requires written justification to continue) |
| Always | Average likes | Less than 0.5 for 2 consecutive weeks | 72h cooling-off + mini-autopsy |
| Always | Weekly operation time | Over 105 minutes for 2 consecutive weeks | Scale down operations |
'Deciding on withdrawal criteria' does not mean 'admitting defeat'.
It is actually the opposite; without criteria, you end up like me,
dragging out life support and wasting 2 months and 24,800 yen.
A system created to crush this failure
For the fourth attempt, I didn't just 'be careful,' I changed the system so I couldn't fail.
I divided strategy, production, and censorship into three AI departments, and before posting, they automatically check for link release dates, posting frequency, hook duplication, and whether I'm writing about things I haven't experienced.
All a human has to do is look at the approval card that arrives on Discord and decide whether to publish or stop it.
I won't hide the results either.
I will publish followers, impressions, and sales every Sunday to verify whether this design actually works.
Once I have results, I will write another article with the full blueprint, including the composition of the three AI departments and the operational rules.
This free article is a 'record of failure,' and the next article will be 'how to prevent the same failure with a system.'
Conclusion
What I learned after three consecutive losses was that,
if failure analysis is just a 'reflection meeting done with willpower,' you will definitely go easy on yourself.
All the logs are left on the disk.
So, let the AI do the reading, and the human just needs to receive the verdict.
Before I create the fourth one, I'm starting by sticking this constitution on my refrigerator.
(I feel like sticking it up physically is the most effective way.)
I hope this serves as a memorial and a reference for others stuck in the same swamp.
If you have any questions, feel free to leave them in the comments.
Is there anyone who wants an autopsy?