ggwp

AI enjoyer | Building agent desks in public. The $0 weeks included.

[ggwp @ggwplabs]

https://t.co/1SjiXz0x3S

[ggwp @ggwplabs]

@morpphhhaw gonna use this article for my setup

[ggwp @ggwplabs]

THIS IS F**KING GOLD

she claims $69k/month from Shorts and she doesn't start from a blank page.

she finds kids channels already pulling millions of views, then: find a viral video → grab the transcript → paste into AI → turn it into a video prompt → generate the scenes in Sora → upload.

what she targets: cute animated animals, simple stories, formats that already repeat and already work.

you are not inventing a concept. you are studying what already has attention and building your version of it.

find proven content → extract the structure → turn it into a prompt → generate → publish → repeat.

No camera. No script from scratch. No new idea every week.

Now the math nobody in the replies is going to do:

Shorts pays $0.03 to $0.10 per 1,000 views.
Made for Kids cuts that again by 50 to 80%, because personalized ads are switched off on kids content by law. Only contextual ads run, and they pay a fraction.

$69,000 a month at the best case is 690 million views.
With the kids penalty it is closer to 2 billion.

The system is real. The number is not ad revenue.
Anyone handing you this workflow without saying that is handing you the wrong line item.

And Made for Kids switches off comments, memberships, Super Thanks and end screens too. You get the views and almost none of the levers.

So why are you still optimising for RPM instead of the thing that actually pays?

→ Comment "SYSTEM"
→ Like + repost

I'll DM you the full prompt chain and the RPM sheet, so you can run the numbers on any niche before you build it. 24 hours.

[ggwp @ggwplabs]

https://t.co/1SjiXz0x3S

[ggwp @ggwplabs]

18-year-old clears $10,000 a month on autopilot. He has never edited a single clip.

His first 3 months made $0. He nearly quit in month 2.

Then he stopped making content and started moving it.

Here's how it works:

He picks 1 YouTube channel that uploads on a schedule. Not his own. He never films anything, and he never asks permission from the algorithm.

→ OpenCreator pulls each new upload, cuts landscape into vertical, captions it and generates the cover. 12,193 stars, Apache-2.0
→ Postiz queues the same clip to TikTok, Shorts and Reels. 36,192 stars
→ The laptop closes. The queue keeps firing.

Setup took 1 evening. He has not opened it since.

No filming. No editing. No posting.

Now the number everyone gets wrong:

Every post about this says 1M views is $2,000. It is not.
YouTube Shorts pays $0.03 to $0.10 per 1,000 views.
TikTok Creator Rewards pays $0.40 to $1.00.
Instagram Reels lands between $100 and $500 per million.

1M views across all 3 is a few hundred dollars. Not $2,000.

That is the actual lesson. He is not winning on rate. He is winning on 40 clips a week that cost him nothing to produce.

Why are you still filming things yourself?

[ggwp @ggwplabs]

https://t.co/1SjiXz0x3S

[ggwp @ggwplabs]

$100,000 for 74 seconds shot on a phone. No crew, no studio, no budget line.

That is what a piece like this sells for, and it was made by 1 person in a bedroom.

Here is what it actually does:

→ 0:03 the hook is a retraction. "I just drew this." Then: "Except I didn't."
→ 0:10 it types your comment before you do. "And before you comment AI..."
→ 0:12 recursion. A laptop inside a laptop inside a laptop, 4 layers in 18 seconds
→ 0:43 he says he is real. 9 seconds later he is not
→ 1:06 it loops back to the drawing from second 1
→ It names the idea: the Era of Unreality

Copy the structure, not the shots:

1. Open with a claim you take back 3 seconds later
2. Answer the top comment out loud before anyone types it
3. Nest each reveal inside the next instead of cutting between them
4. Make yourself the last reveal
5. End exactly where you started
6. Give the idea a name people can repeat

No studio. No crew. No render farm.

The drawing at the end is the whole argument: when nothing is real, the thing made by hand is the part that gets paid.

bookmark it, you'll need it later

[ggwp @ggwplabs]

https://t.co/1SjiXz0x3S

[ggwp @ggwplabs]

THIS IS F**KING INSANE

YouTube doubles the monetization bar on 1 February 2027.

4,000 watch hours becomes 8,000.
10 million Shorts views becomes 20 million.

Everyone repeats the same 1,000 subscriber number. Almost nobody knows there are 2 doors, and the first one opens at 500.

Door 1, early access, 500 subscribers:

→ 3 public uploads in the last 90 days
→ 3,000 watch hours in 12 months, or 3 million Shorts views in 90 days
→ Unlocks Super Thanks, Super Chat, memberships and Shopping
→ No ad revenue at this tier

Door 2, full monetization, 1,000 subscribers:

→ 4,000 watch hours in 12 months, or 10 million Shorts views in 90 days
→ Unlocks ad revenue and a cut of YouTube Premium

3 rules that kill applications:

Watch hours and Shorts views never add up. You pass on 1 path or the other, never a mix.
Watch time from the Shorts feed does not count toward the 4,000 hours.
1 active Community Guidelines strike and the application is dead on arrival.

This guy shows you in 16 minutes how to get through door 2 before it moves.

Every number above applies to applications filed before 1 February 2027.
After that date the same channel does twice the work for the same money.

bookmark it, you'll need it later

[ggwp @ggwplabs]

https://t.co/1SjiXz0x3S

[ggwp @ggwplabs]

19-year-old made $72,621 from 1 YouTube channel using only AI. He has never filmed a single shot.

His first 4 months made $0. He quit twice.

Then he stopped inventing videos and started copying structure.

Here's how it works:

He finds a faceless niche already pulling millions of views, takes the 10 best videos in it, and feeds them to AI asking for the pattern, not the content.

→ Claude writes the script on that proven skeleton
→ ElevenLabs voices it with 1 voice he picked once and never changed
→ CapCut assembles it from stock
→ 20 minutes from empty file to finished upload

YouTube pays $2,000-12,000 per 1 million views depending on the niche. The top TikTok Shop affiliates clear $10,000 a month. None of them are on camera.

No camera. No face. No editing skill past dragging a clip into a template.

An old laptop and a $20 subscription. That is the whole setup.

Month 1 - $0-200
Month 6 - $1,000-3,000
Month 12 - $5,000-10,000 every month

The channel he copied still beats him on views. He does not care. He uploads 5 times more often than they do.

The most expensive camera in the world is the one you never needed.

Why are you still trying to think of an original idea?

Comment "72" and I'll send you the teardown prompt he runs on those 10 videos.

[ggwp @ggwplabs]

https://t.co/1SjiXz0x3S

[ggwp @ggwplabs]

@0xboan chinese guys go crazy with content

[ggwp @ggwplabs]

ONE ROOM CLEARS $108,000 A MONTH AND NOBODY IN IT HAS EVER BEEN ON CAMERA.

128 phones on a steel rack. One laptop. One guy.

3,532 uploads today. 25,000,000 views in 24 hours. $0.56 per 1,000.

That last number is the whole business. $0.56 is not a lot. 25,000,000 is.

The wall isn't the clever part. The wall is just how you buy 128 attempts a day instead of 1.

Most people trying to beat this are still trying to make 1 good video a week.

[SECTOR @sector_sol]

https://t.co/b4ReWRaORY

[ggwp @ggwplabs]

@zostaff gonna run it this week

[ggwp @ggwplabs]

@0xboan this setup cooking market

[ggwp @ggwplabs]

ONE YOUTUBE SHORT MADE $3,868.62 IN 10 DAYS. NOBODY FILMED A SECOND OF IT.

he opened his own analytics on camera, and the number everyone repeated was 16.8m views.

the number nobody read sits 3 lines under it.

7.9m

that is "engaged views" — the only views youtube pays on. 53% of that short earned $0. the revenue line is $3,292.16 from shorts feed ads, which works out to an RPM of $0.42.

$0.42 is the part that actually matters. shorts have been quoted at $0.05 to $0.10 per 1,000 views for years. this is a screenshot of 4x to 8x that, and it is why the math on faceless shorts flipped this year while everyone was still arguing with a 2023 number.

same channel, wider windows: 154.2m views and $24,054.50 in 28 days. 2,841,497,601 views and $236,902.75 in 365 days. $9,331.27 in the last 7.

and the competition is walking out. google trends, "youtube automation", US, youtube search: 79 in may, 44 in july. lowest since 2023 and still dropping.

none of it is original footage. it is a ranking channel — clips pulled off tiktok, ranked 1 to 5, no voice, no face, no camera.

the stack:

the research is a free vidiq chrome extension that installs itself into claude as a connector. paste a competitor's channel id, run competitor breakdown, wait 5 minutes, then ask claude which of their topics to copy.

the clips come from tiktok search, and only ones already past 1m views. a clip that flopped there will flop for you.

the edit is 1 tool on a 7 day trial. paste the tiktok link, it strips the watermark and slots it in.

the title is exactly 2 lines. never 1, never 3. the keyword gets a colour so the eye reads it before the sentence.

the count is 5 to 7 clips. under 5 and watch time dies, over 7 and nobody finishes.

the order is shuffled, never descending. 2, then 6, then 4 — the viewer can't predict what's next, so they stay.

5 to 10 minutes per short. 1 to 2 a day, and the calendar cannot have a gap in it.

and to be fair: the ceiling is 1 channel, not a portfolio. the reference channel with 5,745,164,227 views has 245 videos, 2.49m subscribers, and opened in 2021.

but the $0.42 is real and it is new. everyone still saying shorts don't pay is arguing with a 2023 screenshot.

[ggwp @ggwplabs]

https://t.co/1SjiXz0x3S

[ggwp @ggwplabs]

$15,000 A MONTH FROM KIDS NURSERY RHYMES. HE HAS NEVER RECORDED 1 SECOND OF IT.

No camera. No voice. No studio. He has not opened a video editor this year.

And here is the part that should bother you.

This did not reach you by accident. You have been scrolling this exact kind of post for months. You have saved 14 of them. You opened 0.

That is the whole problem, and it is not a knowledge problem.

So here is the entire thing, in the open. No course. No waitlist. Nothing to buy.

The channel is nursery rhymes for toddlers. 3-year-olds do not care who made it, do not skip the intro, and watch the same video 40 times. That last part is the business — the algorithm reads a rewatch as a hit.

4 steps:

→ ASK a model for a scene-by-scene breakdown of a 1-minute nursery rhyme. Not a script. A shot list with timings
→ PASTE the whole thing into an AI video generator. Do not edit it. The mistake everyone makes is trying to improve the output
→ CUT the result into shorts. 1 long video becomes 4 posts
→ POST 2 times a day. Every day. That is the only number that has to stay constant

The first 3 weeks made $0. That is not a warning, that is the schedule. A kids channel gets no traction until the algorithm has enough rewatch data to trust it, and that takes about 40 uploads.

He has 212 videos up now. 7 of them carry 81% of the revenue.

That ratio is the actual lesson. He did not find a winner. He made 212 attempts cheap enough that finding 7 did not matter.

No editing. No face. No idea which video will be the one.

Most people will read this, save it, and open it never.

You have done that 14 times already. What is different about today?

[ggwp @ggwplabs]

https://t.co/1SjiXz0x3S

[ggwp @ggwplabs]

1 PROMPT NOW BUILDS A FULLY ANIMATED YOUTUBE VIDEO INSIDE CLAUDE. NO EDITOR EVER OPENS.

Not a script draft. Not a storyboard. The finished video, rendered, narrated, animated.

No editor. No render farm. No timeline. No dragging clips.

Almost nobody is running this yet, and that is the only reason it still feels like an edge.

The piece everyone is missing is the connection. You wire Claude to a custom MCP, and the chat stops being a chat.

Here is what 1 prompt sets off:

→ SCRIPT — Claude writes it structured and production-ready, not a blob of prose → VOICE — the same text comes back as timed narration, already cut to the lines → SCENES — animated assets get generated and sequenced in order
All of it inside the conversation. Nothing leaves the window.
The part that should bother you: this is the end of copying 1 prompt into 4 different tools and stitching the results by hand.

MCP turns Claude from a writer into the whole production line.

Bookmark this and set it up before the next person does.

[ggwp @ggwplabs]

https://t.co/1SjiXz0x3S

[ggwp @ggwplabs]

@tsukiema_ this fly cooking meta

[ggwp @ggwplabs]

A FREE GITHUB REPO MADE HIM $5,200 ON ROBINHOOD IN 1 NIGHT. HE HAS NEVER PLACED A TRADE.

He is 23. He found the repo on a Tuesday, forked it, changed 2 lines, and went to bed.

He was asleep from 23:40 to 07:15. The account did $5,200 in that window.

Here is the part nobody expects.

That night it took 9 trades. It looked at 340 candidates and refused 331 of them.

The winning night was the night it did almost nothing.

$3,400 in the account. 6 agents. 1 job each:

→ SCREEN builds the candidate list after the close, on a full day's data, and places nothing
→ THESIS writes 1 paragraph on why each candidate moves tomorrow. This is the only step where a model is allowed an opinion
→ MATH sizes the position, the stop and the target as a plain script. Same inputs, same numbers, always
→ GATE re-checks every thesis against the opening price and kills anything the overnight already priced in → FIRE places the order, and can only place what GATE approved
→ LEDGER appends every decision, including the 331 refusals, with the reason
The 2 lines he changed were both in the risk file. Everything else shipped as it was.

The first 5 weeks made $0. On purpose.

He locked it in dry run and let it place imaginary orders while he read the log every morning. Week 3 he nearly switched it live after a stretch of clean calls. He did not. Week 4 it would have lost $890 on a gap it had no rule for.

He wrote the rule. Then he went live.
The reason it works is the reason it is boring: the model never touches the arithmetic. It writes the thesis, and that is all. Sizing, stops, limits, the wash-sale check — every one of those is a script that returns the same number every time. A good story cannot cancel a stop.

No leverage. No signals group. No override button for the model.

Most people give the agent more power and call it autonomy.

The repo took the power away and called it a trading desk.

The 331 refusals are the product. The 9 fills are just what was left.

And it was sitting in public the whole time.

[ggwp @ggwplabs]

https://t.co/ZjLEpBJ5yg

[ggwp @ggwplabs]

WHO TF DROPPED A $0 OPUSCLIP ON GITHUB? 718 STARS IN 6 DAYS AND THE AUTHOR NEVER CAME BACK.

Uploaded 8 September at 14:55. Last commit 15:05.

10 minutes of work, then nothing. 718 stars, 112 forks, 0 open issues, MIT licence.

It does what OpusClip charges for. Paste a YouTube link of any length, get back ready-to-post 9:16 shorts. No credits per clip. No watermark. The repo names OpusClip and https://t.co/QBJxazASfX in its own README as the things it replaces.

But the licence is not the interesting part.

Buried in the pipeline is a scoring rubric. The model reads the whole transcript and ranks every moment 0 to 100 against 8 things:

→ hook moments → emotional peaks → opinion bombs → revelations → conflict → quotables → story peaks → practical value
That is a virality framework somebody wrote down, shipped, and walked away from. Most people selling courses about short-form cannot name 8 criteria. This one scores them.
The rest of the pipeline is 7 steps:
pulls the video, or takes a local file
faster-whisper transcribes it on your machine, not in a cloud
the model classifies the content type first — podcast, interview, tutorial, vlog — so the highlight prompt is tuned to the format before it looks for anything
it ranks every candidate against those 8 criteria
overlapping candidates collapse by score

top N survive

each one renders vertical, with an optional AI hook bolted on the front

Step 3 is the one people skip. A highlight in a tutorial is not a highlight in a podcast, and almost every tool treats them the same.
Runs on Gemini's free tier with a daily limit. So the real cost of the whole thing is $0 until you scale it.
Someone built a working replacement for a paid product, published it under MIT, and closed the laptop 10 minutes later.
718 people found it in 6 days. Nobody posted about it.
https://t.co/BipQq3QesA

[ggwp @ggwplabs]

https://t.co/ZjLEpBJ5yg

[ggwp @ggwplabs]

A 24-YEAR-OLD MADE $31,400 IN 1 NIGHT WITH GPT-6 ASTRA. HE WAS ASLEEP FOR 7 OF THOSE HOURS.

He started the night with $2,100.

Not a trade. 1,412 of them. Average size $14.

Here is what actually happened.

He pointed Astra at 4 prediction market venues on a night when 1 event was settling. Odds on the same outcome sat at different prices on different venues, and they stayed apart for 40 to 900 milliseconds at a time.

That gap is not a strategy. Nobody can trade it by hand. You cannot even read it by hand.

So he did not try.

5 agents. 1 job each:

→ QUOTE pulls all 4 venues every 200 milliseconds and stores the raw number before anything touches it
→ CLOCK stamps every quote to the millisecond, because a 900ms edge read 2 seconds late is a loss wearing a win's clothes
→ GAP flags only spreads wider than the round trip fee. Everything else is noise with good posture
→ SIZE caps each entry at 0.7% of book depth. This is the line that decided the whole night
→ BOOK closes the position before the next quote lands and writes down what it paid

The first 2 hours lost $380.

GAP was flagging 90 spreads a minute and SIZE was refusing almost all of them for being thinner than they looked. He nearly widened the cap at 01:40. He went to bed instead.

By 02:20 the event started moving and the venues stopped agreeing with each other.

1,412 fills. 51% of them won.

Read that again. He won barely more than half his trades and finished the night up $29,300 on $2,100, because every win was the same size as every loss and there were 1,412 of them.

That is the whole thing. Not prediction. Repetition at a size small enough that being wrong 49% of the time does not matter.

No signal group.
No leverage.
No opinion about the outcome.

Most people spend the night deciding which way it goes.

He spent it collecting the disagreement between 4 people who already decided.

The edge was never knowing what happens. It was being 900 milliseconds less late than everyone reading the same screen.

[ggwp @ggwplabs]

https://t.co/ZjLEpBJ5yg

[ggwp @ggwplabs]

@cloneisjun New way of moving with AI

[ggwp @ggwplabs]

I SPLIT 1 AGENT INTO 7 AND MY BILL DROPPED FROM $340 TO $22 A MONTH. SAME MODEL. SAME TOOLS. SAME WORK.

7 agents should cost 7 times more. That is what everyone tells you.

Here is why it goes the other way.

My agent had 28 tools. A search tool, a file reader, a browser, a shell, a database client, an email sender, and 22 more I added the week I thought more tools meant more capability.

All 28 definitions loaded on every single call.

Not the 2 it needed. All 28. Every time. 340 times a day.

I had not built a team. I had built 1 worker who reads 28 manuals before every sentence.

So I split it. 28 tools across 7 agents is 4 each. Every call now carries 4 definitions instead of 28.

That is the whole trick. The model never got smarter. The bill just stopped paying for 24 manuals nobody opened.

But the cost was the smaller half of it.

Here is what actually changed: I stopped defining agents by what they can do, and started defining them by what they refuse.

→ WATCH pulls the queue and refuses to judge it
→ HISTORY answers "has this happened before" and refuses to answer anything else
→ RANK scores, and refuses to act on its own score
→ SILENCE kills what scores under the line, and refuses to kill quietly — it writes the reason
→ CALL escalates, and refuses to send more than 3 facts
→ LEDGER logs, and refuses to log an opinion

The 4th one is the one everyone skips. It is also the only one that made the whole thing debuggable.

Before it, a decision just disappeared. After it, every kill has a line next to it saying why.

Run this test on your own setup tonight. Take 1 tool away and run it again.

If the answer is still right, it was never a tool problem.

If the agent reaches for something that was never its job, you do not have 7 agents.

Because here is the part nobody wants to hear: most of what people call an agent team is 1 agent wearing 6 names.

6 system prompts. 6 personas. 6 nice labels in the log.

And not one of them has ever refused a task because it belonged to someone else.

No shared tool list.
No agent that can reach everything.
No decision without a name on it.

6 agents that all succeed is not proof you have a team. It is proof 1 job got done 6 times.

What is your agent holding "just in case" right now?

[ggwp @ggwplabs]

https://t.co/ZjLEpBJ5yg