How a video gets made here

10 stages · 11 skills · one folder per video · one human who clicks publish

The problem the system solves is not quality

0→1

I build things easily.
Then I freeze at launch. Apps built, never shipped. Domains bought, never started.

So every stage is designed to make publishing cheaper than not publishing.

The box: start a timer, publish at 3:00 whatever exists

Block	What	Time box
1	Idea, research, packaging, beat map, slides	60-90 min
2	Record, one take	30-40 min
3	Edit, captions, copy, thumbnails (mostly unattended)	35-45 min
4	Publish (manual)	15 min

A published 6/10 beats an unpublished 9/10. Fixing is allowed. Not shipping is not.

Ten stages, cyan = AI drafts, violet = human only

One folder per video is the whole memory

Before recording

brief.md research.md titles.md
beats.md talking-points.html
slides/deck.html thumbnail/

After recording

recording.mp4 final.mp4 final-loud.mp4
final.srt description.md posts.md
substack.md review.md

Every stage writes files, never chat. A fresh AI session can pick up any stage cold.

STAGE 1

Ideas: capture first, validate second

Phone idea → appended to the backlog in my exact words, lossless, then tidied
Five pillars, one number per idea, privacy flag inline, never dropped
Before it becomes a video: do similar videos show a live audience?

Frustrations in the daily log are ideas too. The log is mined newest first.

STAGE 2

Research: every number I say exists in a file

3-5 angles, web fan-out in parallel, 3-5 strongest sources each
Each claim labeled [V]erified · [E]xtracted · [D]irectional
An "honesty guardrails" section: what NOT to oversell
Sources saved to Reader, tagged by video number, archived after publish

STAGE 3

Packaging before the outline

20

title candidates, then 3 picks A, B, C
each paired to one of 3 thumbnail drafts, armed for YouTube's A/B test

If no title excites, the idea is not ready. It goes back to the backlog.

STAGE 4

Beat map: open loops, close loops

#	Beat	Opens	Closes	Mini-hook out	Slide
1	Hook + stake	2-3 loops		"so the first rule…"	face
…	≤ 5 beats total		smallest first	one line each	id

The hook is the only fully scripted text. Outline makes you talk. Script makes you read.

STAGE 5

Slides: built, audited, never presented from source

HTML + Mermaid, offline, slide IDs match the beat map 1:1
≤ 5 lines of text per slide, numbers big, diagrams over bullets
Bottom sixth of the frame is footer only: captions live there
Every slide screenshotted at 1920×1080 and looked at before handoff

STAGE 6

Record: the AI writes the guide, then goes silent

Pre-flight: scene, mic peaks yellow, camera on wall power, DND, privacy shot-list
Guide is one HTML page on the other monitor. Hook verbatim is item 1
One take. Flub = pause 2 s, clap once, restart the sentence
Hook re-recorded last if it did not land

STAGE 7

Edit: one command, mostly unattended

Review rule: watch the first 60 s and the last 30 s. Do not rewatch the whole thing.

STAGE 8

Copy: only from what was said

Transcript → chapters phrased as curiosity, description, posts, pin comment
Thumbnails rebuilt from a still of THIS footage, 168 px shrink test
Privacy frame-check at every screen-share chapter
Report card on 8 basics: intro, title, thumb, story, jargon, cuts, audio

The report card never blocks a publish.

STAGE 9

Publish is manual, in this order

Nothing goes public without a human reading it. Unlisted → public: two small clicks beat one big one.

STAGE 10

Retro: designed, never ran

0

48-hour retro files written, out of 37 videos
the skill exists, the loop never closed

What ran instead: the report card, 24 times, tracking craft strikes video over video.

Who does what

AI drafts

research, titles, beats, slides, guide
cuts, captions, spelling, loudness
chapters, description, posts, article
thumbnails, privacy frame scans

Human decides

the idea of the day, the take
ear-check final vs final-loud
title + thumbnail pick, 15 min cap
every click that makes it public

What actually happened: cadence

30

days, a video every day, July 20 to Aug 18
one missed day, two published the next. Two dates had to be repaired after the fact.

7

videos in the 16 days after
non-daily by decision. One fully prepped video was never recorded.

Plan vs improvise

~9

of 37 videos came from the planned backlog
the rest: same-day news, sequels, or whatever the benchmark rig finished overnight

"Packaging locked before the script" held about two thirds of the time. The footage re-opened it the other third, because the video did not say what the thumbnail promised.

Hold that thought: the improvised videos are the ones that worked.

The 2-3 hour budget: never clocked

What is recorded	Example
raw → final runtime	33:24 → 28:09 (car math) · 9:15 → 8:51 (cleanest take)
record start times	23:34 · 23:56 · 00:59 · 00:31 · 23:12
total session time	not once

The one metric the plan said it would measure was the one never instrumented.

Where the AI failed or got overruled

Audio mastering: better numbers, "sounds worse." Five variants rejected by ear
Screen-share leaks in 12 videos: logins, history dropdowns, a plaintext password
Mis-heard names every video: 80 fixes on one, one changed a dollar figure
13+ thumbnail rejects, text over face, half-closed eyes, too much me
Its own wrong fact once: called a price 3× when it was $100 more

Where it clearly won

Overnight fact-check cut two of my own wrong on-camera claims before publish
Four nights in a row: record → edit → thumbs → Substack → YouTube → site, same night
Eight data tables generated from results.json, zero hand-copying
A second address-bar leak found by a color scan my eyes missed
Read a model's chat template out of a GGUF binary to settle an argument

Every failure became a rule

Date	Rule added
Jul 25	3 titles × 3 thumbnails, always
Jul 28	A/V sync 0.26 s, clap in frame
Aug 06	Hook first, greeting second
Aug 08	Substack first · links as text · raw archived before edit
Aug 14	Ear test outranks metrics
Aug 25	Caption spelling dictionary, automatic

Still manual, 46 days in

Every publish click, on three platforms
Substack: no API. Render, open, select, paste. Tables flatten, so they ship as images
Thumbnail final pick, and the ear check on two audio versions
Analytics: no API key, a Studio CSV export by hand. Tonight was the first one
"Welcome back" still opens videos: flagged 15 times

The number I never looked at, until tonight

642

views, 9 subscribers, on Aug 8
the only channel datapoint in 46 days of logs

76,551

views, 462 subscribers, 4,113 hours watched, on Sep 3
YouTube Studio export, pulled tonight. 617,719 impressions, 5.08% click-through.

Engaged views per day, Jul 20 to Sep 2

Six videos are 87% of every view

Video	Views	CTR	Avg viewed	Subs
Qwen 3.8-Flash-Next on 128 GB: 1-bit to 4-bi…	17,803	6.5%	20.03%	42
I tested every thinking level on Qwen 3.8-27B	13,041	4.7%	17.7%	142
Why I stopped buying $400 keyboards for a ti…	10,700	5.49%	20.47%	22
I shrank Qwen 3.8-27B to 11GB	9,877	6.39%	16.99%	107
Qwen 3.8-27B vs 3.6-35B: I made them fight	7,479	4.13%	29.04%	44
Qwen 3.8-Flash-Next just dropped. Can I even…	7,451	6.75%	22.1%	10

Five of the six are local-AI benchmarks that started as same-day pivots. The sixth is a keyboard review.

What the numbers say

83%

of all views and 410 of 462 subscribers came from 11 benchmark and local-AI videos
none of them were in the original plan

~1,400

views across the first 21 daily videos combined
median video on the channel: 72 views. 19 of 36 are under 100.

Three takeaways

Design for the freeze, not for quality. The streak was the product
Files over chat. Every AI stage writes into the folder, so any session can resume
The AI drafts everything and decides nothing public. Every failure above became a rule
The plan got the streak going. The pivots got the views. Look at the numbers sooner than 46 days in

What did you build and never launch?

Tell me in the comments. I will tell you which stage would have unfrozen it.

Every skill file and the edit chain are described step by step at hijoseromero.com