Handbook · by Dr. Louis Korczowski

The Content Testing Loop: make every short video teach the next one, with AI reading your analytics.

Alex Hormozi spent $4 million over about 40 months on team, studios, vendors and software, and learned by comparing his best and worst videos. A founder can run the same loop in about two hours a week: one table, a seven-day rule, a weekly post-mortem, a format review and a monthly learning session. Every recipe ends with one thing to do today and a prompt for your assistant.

Ask your AISummarize this handbook and apply it to your situation: ChatGPT Claude Perplexity Gemini Markdown AI skills on GitHub

TL;DR

A weekly routine that makes every short video teach the next one. Log each video in one table with its format and hook. Wait seven days before any verdict. Once a week, compare your most surprising video with ten of your own comparable videos, passage by passage, and re-edit it with one to three changes. In the same weekly session, review your formats as dice: keep, change one variable, or drop. Once a month, re-test your saved lessons and add new ones only with examples and counterexamples. An AI assistant does the reading in minutes. You keep the ideas and the decisions.

The method

Each video teaches the next one, on your own data.

Eight recipes. Each one fits on a phone screen, ends with one thing to do today, and comes with a prompt for your AI assistant.

Fig. 1 · The content testing loop● the post-mortem is where the AI saves you the most time

When	What	Time
Every video	Log it, tag format and hook	2 min
Day 7 of each video	Read the numbers, mark it mature	1 min
Once a week	Post-mortem, re-edit, test, dice review	75 min
Once a month	Learning session and lessons ledger	60 min

01Log every video in one table30 min once · then 2 min per video
02Tag each video with its format and its hook1 min per video
03Wait for mature data before any verdict7 days · every video
04Run one post-mortem a week20 min · weekly
05Re-edit with KEEP, CHANGE and CUT15 min · weekly
06Test two versions fairly10 min · per test
07Review your formats as dice30 min · weekly
08Hold a monthly learning session60 min · monthly

The recipes

Eight recipes to run the loop every week.

Log every video in one table

30 min once · then 2 min per video

One row per video, read on the same day of its life. Without a table, an archive is a pile of memories.

You need

A spreadsheet (Google Sheets, Excel, Airtable)
The analytics screen of your platform
The transcript or script of each video

Steps

Create the columns: video id, platform, publish date, reading date, days live, distribution (main feed, test, ad), format, hook type, length, views, likes, comments, saves, shares, new followers, average watch time, retention at 3 s, 10 s and the end.
Read the numbers on day 7 for every video, and write the reading date. Day 7 is the reference line.
Keep the transcript of each video in a folder named by video id.
Optional: copy the retention curve second by second for the videos you will study (one row per second, with what is said and shown).

Platforms give creators these numbers per video: TikTok's analytics show views, average watch time, the share who watched to the end, traffic sources and the moment viewers lost interest (TikTok Creator Academy). TikTok Studio on desktop exports a CSV. Copying the numbers by hand on one fixed day stays clear of TikTok's ban on automated extraction: its terms forbid extracting data with automated software unless TikTok approves it in writing (TikTok Terms of Service).

Avoid

Comparing a 2-day-old video with a 3-month-old one
Mixing paid and organic views in the same column
Typing numbers from memory

TodayCreate the table and fill the rows of your last ten videos.

For your AI

Copy this prompt into your assistant:

Here are the analytics of my last 10 videos: [paste or screenshots]. Build my video log with these columns: video id, platform, publish date, reading date, days live, distribution, format, hook type, length, views, likes, comments, saves, shares, new followers, average watch time, retention at 3 s, 10 s and end. Flag every missing value.

Skill: video-post-mortem · install: npx skills add lkorczowski/content-strategist-AI-cookbook --skill video-post-mortem

Tag each video with its format and its hook

1 min per video

You learn about formats. A single video teaches very little. The tag is what turns a list into evidence.

Steps

Name your formats with short, stable names: founder story, myth vs data, reaction, teardown, tier list.
Tag the hook type of each video: a claim, a question, a number, a contradiction, a confession, a piece of news.
Write the first on-screen words and the first spoken sentence in the table.
Never rename a format in the middle of a test. Create a new one instead.

A format is a repeatable shape: same opening, same structure, same length, a new topic each time. Judging the shape across many videos is how luck gets separated from craft, the core of the dice system described on the method page.

TodayTag the ten videos already in your table.

For your AI

Copy this prompt into your assistant:

Here are my last 20 videos with their transcripts: [paste]. Propose 4 to 8 stable format names and a hook type for each video. Explain each tag in one line.

Skill: format-dice-review · install: npx skills add lkorczowski/content-strategist-AI-cookbook --skill format-dice-review

Wait for mature data before any verdict

7 days · every video

Before day 7 and 1,000 views, describe the video and keep the verdict for later.

Steps

Mark every video younger than 7 days or under 1,000 views as early data. Use your usual floor if your account is smaller.
Early data can be described ("held better than usual at 3 s"). It cannot be explained.
Re-read the numbers on day 7 and write the new reading date.
Keep the rule even when a video looks like a hit on day 1.

Views are very unevenly spread: in one large YouTube study the top 10% of videos drew nearly 80% of views (Cha et al., 2009). Early numbers move a lot, and checking a test again and again until it looks good fools you: in Evan Miller's worked example, testing after every observation and stopping at the first significant result pushed false positives from 5% to 26% (Evan Miller, How Not To Run an A/B Test). The 7-day and 1,000-view thresholds are a working rule shared publicly by Angus Sewell: adapt them to the size of your account.

Avoid

Rewriting a format after one bad day
Reading a story into a curve measured on 300 views
Checking the numbers every hour

TodayWrite the maturity rule at the top of your table.

For your AI

Copy this prompt into your assistant:

In my video log [paste], mark each video as mature or early data (mature = 7 days and 1,000 views). Describe the early ones in one line each, without any verdict.

Skill: video-post-mortem · install: npx skills add lkorczowski/content-strategist-AI-cookbook --skill video-post-mortem

Run one post-mortem a week

20 min · weekly

Pick your most surprising mature video, the best or the worst, and compare it with your own comparable videos, passage by passage.

You need

The video log
The transcript and on-screen text of the video
Its retention curve, or at least 3 s, 10 s and end
10 to 20 comparable videos (same platform, same format, similar length)

Steps

Cut the video into windows: 0 to 2 s, 2 to 5 s, then 2 to 3 seconds aligned on sentences.
For each window, note the percent still watching, what is said, what is on screen and the drop.
Compare each drop with the usual drop of your comparable videos at the same moment. Judge the first 2 seconds against your other first 2 seconds only.
Name the worst reliable passage and the best one, and read the context: a promise not paid, a slow setup, a jump in topic, on-screen text that contradicts the voice.
Check each explanation on two other videos of your archive, and look for videos that contradict it. If you cannot find them, write the word hunch.

Alex Hormozi described the hand-made version of this recipe: he compared his 12 weakest and 12 strongest recent videos to see what the winners had in common (The Game, Ep. 750, 2024). An AI assistant now does the same reading in minutes, passage by passage. Your own archive is the yardstick because your audience, topics and format set your normal curve.

Avoid

Comparing with industry averages
Merging retention, saves and shares into one score
Changing ten things at once

TodayChoose the video for this week's post-mortem.

For your AI

Copy this prompt into your assistant:

Run a video post-mortem on video [id] against these comparable videos: [paste log rows, transcripts, retention]. Give me the passage table, the worst reliable passage and the best one, each explanation marked supported or hunch, and what the data cannot tell.

Skill: video-post-mortem · install: npx skills add lkorczowski/content-strategist-AI-cookbook --skill video-post-mortem

Re-edit with KEEP, CHANGE and CUT

15 min · weekly

One to three changes, with exact timecodes. Every change is a test until a re-edit beats the original.

Steps

KEEP: the passages that held better than usual. They must survive the re-edit.
CHANGE: rewrite one line or one visual, word for word, where the worst passage starts.
CUT: remove the passage with exact start and end times when it adds nothing.
Label each change as a test, with the metric that will judge it (retention at 10 s, saves, shares).

Changing one variable at a time is the simplest way to know which change worked. The same rule runs every controlled experiment, from a lab bench to an ad split test.

TodayWrite one CHANGE line for this week's video.

For your AI

Copy this prompt into your assistant:

From this post-mortem [paste], write the edit plan: KEEP, CHANGE and CUT with exact timecodes and the new lines word for word. Maximum three changes. Label each one as a test with its deciding metric.

Skill: video-post-mortem · install: npx skills add lkorczowski/content-strategist-AI-cookbook --skill video-post-mortem

Test two versions fairly

10 min · per test

Same day of the week, same hour, same platform, one difference. Decide the winner before you post.

Steps

Choose the deciding metric and the reading date before posting.
On Instagram, post the re-edit as a Trial Reel: it is shown to non-followers first, and you choose to share it with everyone after the results (or let it auto-share when it performs in the first 72 hours).
If the platform has no test feature, post the re-edit at the same hour and weekday as the original, a week later, and compare day-7 numbers.
Promote the winner. Write the result in the table, even when the original wins.

Instagram Trial Reels launched in December 2024 (Instagram) and became available to everyone in June 2025 (Meta, June 2025). TikTok's split testing covers ads only, with at least 7 days and 80% statistical power recommended (TikTok Ads Help). YouTube's Test & Compare picks the winning title or thumbnail by watch time share and excludes Shorts (YouTube Help).

Avoid

Deleting the losing version to hide it
Deciding the metric after seeing the numbers

TodayWrite the deciding metric of your next test.

For your AI

Copy this prompt into your assistant:

I want to test version A [describe] against version B [describe] on [platform]. Tell me how to post them in comparable conditions, which metric decides, and on which day I read the result.

Review your formats as dice

30 min · weekly

A video is a throw. A format is the die. Judge the die on many throws, then keep it, change one variable, or drop it.

Steps

Group mature videos by format. For each format: median views, the typical range, the floor, saves and shares per 1,000 views, retention at 3 s.
KEEP the formats above your median with an acceptable range. Make them templates and produce variations.
CHANGE ONE VARIABLE on formats close to your median or with a wide range: only the hook, the length or the setting.
DROP a format only after the repetitions you planned for it, and after one change. Replace it with a new one.
Keep both kinds of dice: formats that copy what already works on the platform, and formats built around you.

Views are spread very unevenly: most videos land in a band and a few explode. A median and a range describe a format far better than its best video. The theory is in the essay on the dice system.

Avoid

Dropping a format after two videos
Keeping a format because one video went viral
Running only borrowed formats

TodayCount how many mature videos each of your formats has.

For your AI

Copy this prompt into your assistant:

Run a dice review on my video log [paste]: median, range and floor per format, saves and shares per 1,000 views, a KEEP, CHANGE or DROP decision with one reason each, and a plan for my next production day of [n] videos.

Skill: format-dice-review · install: npx skills add lkorczowski/content-strategist-AI-cookbook --skill format-dice-review

Hold a monthly learning session

60 min · monthly

Re-test what you think you know before you learn anything new. A refuted lesson stays in the ledger with the word REFUTED.

You need

The whole video log
A lessons ledger: lesson, scope, status, examples, counterexamples, first seen, last checked, next test

Steps

Split the archive by platform, and by period when something changed (a paid boost, a long pause, a new format).
Re-test every saved lesson on the new videos: confirmed, weakened or refuted.
Look for new patterns by format, hook, first words, length, topic and posting time. Compare medians.
Add a lesson as a hunch when seen once or twice, supported when seen in five videos across two topics, confirmed after one dedicated test.
Choose the next three tests, one variable each.
Ask your assistant for a 300 to 500 word coaching brief, and listen to it as audio if you prefer: Gemini Notebook, formerly NotebookLM, turns a document into an audio overview, and Google warns it can contain inaccuracies.

Keeping refuted lessons in the ledger, with the word REFUTED, stops the team from learning the same false lesson twice. Separate platforms because they weigh signals differently: TikTok treats finishing a longer video as a strong interest signal (TikTok), and Instagram's Adam Mosseri names watch time, likes and sends as the top three signals (Social Media Today, 2025).

Avoid

Keeping lessons that describe one viral video
Deleting refuted lessons
Mixing organic and paid results

TodayCreate the lessons ledger with three columns: lesson, status, evidence.

For your AI

Copy this prompt into your assistant:

Run a content learning session on my video log [paste] and my lessons ledger [paste]. Re-test each saved lesson first, then add new ones with examples and counterexamples, choose the next three tests and write a 400-word coaching brief.

Skill: content-learning-session · install: npx skills add lkorczowski/content-strategist-AI-cookbook --skill content-learning-session

For your AI

The same method, written for your assistant.

Skills in the open Agent Skills format, maintained on GitHub so every update reaches you. Upload them in Claude, paste them in ChatGPT, or install them with one command.

video-post-mortem

Compare a video passage by passage with your own comparable videos and get KEEP, CHANGE or CUT edits.

format-dice-review

Review your formats as dice on mature data: keep, change one variable or drop, then plan the next production day.

content-learning-session

Monthly: re-test saved lessons, add new ones with examples and counterexamples, and get the next three tests.

Rules every skill follows. Drafts only: the AI never sends or publishes for you. No passwords: you sign in yourself. No promised views.

What I commit to

Some videos will do two thousand views, some will reach hundreds of thousands, and a few will pass a million. We run so many repetitions that luck plays a smaller part every month, and I don't drop a format until we have run the repetitions we planned and reached the results we set together.

Credit

This handbook was triggered by a public workflow shared by Angus Sewell (Angus the Nontechnical, TikTok, October 22, 2026), who rebuilt Alex Hormozi's testing logic with an AI assistant. The structure of several recipes comes from his system: the early-data rule (recipe 03), the worst reliable passage (recipe 04), KEEP, CHANGE and CUT with speculative changes labelled as tests (recipe 05), the trial-reel A/B test (recipe 06), and the learning session that re-tests saved lessons first, keeps REFUTED lines in the ledger and ends with a coaching brief (recipe 08). The prompts and skills here are written from scratch, in my own words, and add the format review of the dice system I run with founders (recipe 07).

The theory behind it

Why David Ogilvy would have loved second-by-second retention, what Hormozi's content budget actually buys, and what AI changes for founders. Read the essay by Dr. Louis Korczowski on louis-korczowski.org. The dice system itself is explained on the method page.

Questions

Content testing with AI, answered.

What is a content testing loop?

A weekly routine where every short video is measured against your own comparable videos, one change is tested at a time, and lessons are written down with their evidence, so each video teaches the next one.

How many videos do I need before the loop works?

Start the loop from your first video, but judge a format on at least ten mature videos and a single change on a test planned before posting. With fewer videos, describe what you see and keep the conclusion for later.

Which metric matters most?

Read them apart, because they answer different questions: retention says people stayed, saves say they want to come back, shares say they want someone else to see it. For a founder, shares and the leads they bring usually matter more than raw views.

Do I need to code or automate the analytics?

No. A spreadsheet and ten minutes a week are enough to start. Automate later with the platform's official tools where they exist, read each platform's terms first, and never give an assistant your password.

What does the AI do, and what stays with me?

The assistant reads the curves, compares passages and drafts the edit plan in minutes. The ideas, the decision to keep or drop a format, and the courage to publish stay with you.

Version 1.0 · last updated 2026-10-24: First edition: eight recipes, a prompt per recipe and three companion skills in the public cookbook. Changes are logged in the changelog. Author: Dr. Louis Korczowski.

Bring the three ideas you repeat on every sales call.

Work with me

Two-minute application. I reply within 24 hours.