What does title do?
The title comes before the script. Title and thumbnail set what the viewer expects; the video must deliver exactly that. The Scout writes the title after the idea step and before the proof and the script.
Not here: the thumbnail image (skills/thumbnail/), the first sentence of the video (skills/script/).
Method
- Read the input. From the idea step: topic, tool, planned test, the numbers it will measure, proof videos. Read
brand/rules/title.mdand the title limits inchannel.json(crewcut brand show). If a number is not measured yet, write[N]and mark the title as provisional. - Write five candidates, each with a different hook. A hook is the reason to click: free or money, a time promise,
something new, excitement, an explanation for beginners. The click words that work in the user's language are in
voice.md(## Title hooks) when the Brand identity found them; otherwise judge by the proof videos.- The best-known name early. The thing the title is about is the best-known thing of its class (the famous
product, not the category word), and it starts within the first characters (
title.name_within). Phones cut long titles. - A number only when it is real. It comes from the own test and matches what the video shows.
- A price ladder ("for 1, 10 and 100") or a contrast ("cheap beats expensive") only when the video measures every step or shows the contrast. If the measurement turns out differently, the title changes before the upload.
- Plain words and a promise. No category jargon, no business words in the title: the title speaks to everyone, the business case sits inside the video.
- The best-known name early. The thing the title is about is the best-known thing of its class (the famous
product, not the category word), and it starts within the first characters (
- Check.
python3 skills/title/scripts/title_check.py "Title 1" … "Title 5" --name "<tool>". A candidate with an error does not go to the user; rewrite until all five pass. A hint "No hook" means: sharpen the reason to click. The script ranks with points; the points are a ranking aid, not a forecast. - Thumbnail text. At most
thumbnail.max_wordswords: one key word of the title and the verdict. Never the whole title. Check it with--thumbnail "<text>". - Three variants for the A/B test (
channel.jsontitle.variants): place 1 the recommended title; place 2 the same title with exactly one change (the hypothesis: without the number, or with one click word); place 3 a calmer variant, or for an old video its current title as control. Check all three. - Write
brief/TITLE-vNNN.md: chosen title, characters, the promise, the hook, the number and where it comes from, the thumbnail text, the three test variants, rejected candidates with the reason. Never overwrite a version. - Humanizer pass on the words. Before the card, run the Humanizer skill over the five titles and the thumbnail
text (write them first to
work/titles-draft-vNNN.md), the same way as the script (skills/script, "Humanizer pass"): look for advertising language, inflated meaning and staged run-ups; the click words ofvoice.md(## Title hooks) and every number stay. Record it:crewcut humanize record --draft work/titles-draft-vNNN.md --final brief/TITLE-vNNN.md. The card then shows what changed. - Card.
crewcut card propose --episode <id> --step title --kind title --title "Pick a title" --file brief/TITLE-vNNN.md --body "Recommendation: option 1. <one sentence why>" --option "<1>" … --option "<5>", thencrewcut card wait. - Hand over. To the proof step: the promise and the number to measure. To the script: promise, tool, number. To the Designer: title, thumbnail text, verdict, the three variants.
- After the proof, check again. Do the number and the hook still hold? If not, write a new version and a new card before the script starts.
A title change after the upload needs a card and goes through skills/upload/.
What the check cannot see
Whether the number is real, whether the video keeps the hook, which name is better known, whether every step of a ladder is measured. That is reading work in steps 2 and 9.
Card format
Title (recommendation first)
1. <title> (<N> characters) · hook: <…> · reason: <proof video or rule> <- recommendation
2. …
Thumbnail: <key word + verdict>
All five pass title_check.py (points: 64, 61, …).
Planned
- A results table from the 48-hour checks (
review/results-48h.jsonof past episodes) that the title step reads to weigh hooks by the user's own data. - Language data packs: hook words with measured effects per language, built from public title studies (the studio had a German one; its raw data is not part of Crewcut).
Sources
- Vexian in Grant Owen's podcast, "The Crazy Genius Who Solved The Youtube Algorithm": https://www.youtube.com/watch?v=41HMFGbENj4. Taken: the best-known thing of its class, click words that work in any niche, the price ladder, the contrast that the video must prove.
- MrBeast's production handbook (public since 2024), summary: https://simonwillison.net/2024/Sep/15/how-to-succeed-in-mrbeast-production/ Taken: title and thumbnail are fixed before the shoot, and the video must deliver them.