← AI Creator CourseAll programsHomeSearch
AI Creator Course·AI Content Planning & Writing·1:44:16

Sections 7-8: Thumbnails That Stop the Scroll, and Writing With AI — Train It on Your Voice, Ideas to Outline to Script, Canvas, Research, Feedback, Repurposing, and the 5-Part Ad Formula

Anthony Gallo Instructor - runs the 'Content Creators' YouTube channel and uses its real thumbnails, split tests, scripts and ad results as the examples

The short version

  1. Thumbnail theory in four rules: stop the scroll (instantly recognizable - a person holding the fish, not the lure alone); busy is bad (one primary subject, at most one secondary, one or two colours); no text or four words max; engineer curiosity (a confused face, two mismatched objects, an arrow to a detail, a two-word claim like 'Yikes'). Evidence from his channel: 44.6% vs 29% watch share in a split test; a student's CTR from 0.5% to 4.1%.
  2. Thumbnails in Nano Banana Pro: one good reference (or a character sheet), a plain-language prompt for background, placement, expression, secondary objects and text, 16:9 at 2K, two outputs - then re-upload and edit ('move him closer', 'enlarge the icons', 'add a stroke') instead of chasing one perfect prompt. The Triple Threat GPT gives three routes: from scratch, improve your idea, or take inspiration from another creator's thumbnail - 'there is always a line between inspiration and flat out copying.'
  3. Train the AI on your voice once: 'Can you analyze my writing style? Add it to your memory and in the future write content similar to how I wrote this' plus a standing teleprompter-format instruction; correct word choices as you go; no sample? dictate a story into your phone.
  4. Pipeline: seed a thread with channel description, audience and the channels your audience watches -> 10 curiosity-driven titles -> 5 variations under 60 characters -> ask which is most SEO-optimized -> outline -> edit it by hand -> script section by section because 'ChatGPT is like a muscle... the more we ask in one request, the more it fatigues' -> read it aloud. Canvas (created directly, not via the button) adds rewrite-for-tone, expand, shorten, reading level and Final Polish.
  5. Research with seeded memory: trends, constrained recommendations, comparison tables, opposing-perspective paragraphs to counter your bias, beginner analogies, statistics it will admit it cannot verify, FAQs as a completeness checklist. Feedback: paste the script, ask for factual inaccuracies and style notes; apply by hand ('control freak'). Repurpose one script into a blog (Google traffic; embedding the video is rewarded by YouTube), a shorter newsletter with a CTA, and 3-5 shorts.
  6. Ads: Hook -> Problem/Desire -> Solution -> Offer -> CTA, plus social proof, urgency, guarantee. 'If you do not hook your audience in the first sentence or two, the rest of your ad doesn't even matter.' Three AI strategies: full-AI from a worksheet with the copywriter prompt (save it as a Project - an 'ad writing machine'); lay a foundation with your own hook ('My first AI ad cost $6... $50,000 in the first month'); remix a winner and ask for 3-4 more shocking hooks.

At a glance, three clicks deep

Skim here first: the closed row is the glance, open is the study card with the key points and timestamps, and the ↓ link drops to that concept's full write-up below.

01Thumbnail theory: stop the scroll, busy is bad, four words max, engineer curiosityInstant recognizability;

Instant recognizability; one subject + one helper + two colours; <=4 words; a curiosity trigger - with measured lifts.

'You need to stop the scroll' (p2197590889 0:00)

One primary, at most one secondary subject; 1-2 colours (p2197590889 0:02-0:03)

Zero text ideal; four words max (p2197590889 0:03-0:04)

Curiosity: face, mismatch, arrow, short claim (p2197590889 0:04-0:05)

Split test 44.6% vs 29%; 'Yikes' A/B; CTR 0.5% -> 4.1% (p2197590889 0:08-0:09, 0:19)

↓ Full write-up of this concept

02Thumbnails in Nano Banana Pro, and the Triple Threat GPTReference -> descriptive prompt -> 16:9/2K/x2 -> re-upload and edit;

Reference -> descriptive prompt -> 16:9/2K/x2 -> re-upload and edit; GPT modes: scratch / improve / inspire; inspiration != copying.

Full verbatim thumbnail prompt (p2197590888 0:02-0:03)

16:9, 2K, two outputs (p2197590888 0:03)

Iterate by re-upload + edit-only instructions (p2197590888 0:04)

Multi-reference person + product (p2197590888 0:05-0:06)

Triple Threat GPT three modes; thumbnail-grabber; 'line between inspiration and flat out copying' (p2197590888 0:09-0:16)

↓ Full write-up of this concept

03Train AI to write like you: two prompts and a correction habitStyle-analysis prompt into Memory + standing format instruction + ongoing corrections;

Style-analysis prompt into Memory + standing format instruction + ongoing corrections; voice-to-text as a seed.

Verbatim style prompt (p2186925418 0:00)

Verbatim teleprompter-format prompt (p2186925418 0:01)

Correct word choices and ask it to remember (p2186925418 0:02)

No sample: write one or dictate one (p2186925418 0:02)

↓ Full write-up of this concept

04Ideas -> titles -> outline -> script, one section at a timeContext -> 10 ideas -> 5 title variants <60 chars -> SEO pick -> outline (hand-edited) -> section-by-sectio…

Context -> 10 ideas -> 5 title variants <60 chars -> SEO pick -> outline (hand-edited) -> section-by-section script -> read aloud.

Seed with channel context and 'channels your audience watches' (p2186925426 0:03-0:05)

Verbatim 10-ideas prompt (p2186925426 0:06)

Verbatim 5-variations-under-60-characters prompt (p2186925426 0:10)

Outline prompt; edit before drafting (p2186925429 0:01-0:02)

Section-by-section with build-on-previous instruction (p2186925429 0:03-0:04)

Read aloud (p2186925429 0:04-0:05)

↓ Full write-up of this concept

05ChatGPT Canvas as the writing desk (create it directly, or the tools vanish)Canvas = persistent editable doc with tone/length/level tools;

Canvas = persistent editable doc with tone/length/level tools; create via button or /canvas, not by converting a reply.

Open Canvas, name the doc, edit live (p2186925433 0:01)

Tool list (p2186925433 0:01-0:03)

Workflow prompt (p2186925433 0:04)

Button-converted replies lack the AI tools (p2186925433 0:04-0:05)

↓ Full write-up of this concept

06Research with seeded memory, feedback without surrender, and the repurposing cascadeSeeded research prompts;

Seeded research prompts; script feedback prompt applied manually; script -> blog -> newsletter + CTA -> 3-5 shorts.

Seed memory; trend prompt; constrained recommendations; tables; opposing views (p2186925439 0:00-0:09)

Beginner-analogy prompt; honest stats; FAQ checklist (p2186925439 0:10-0:13)

NoteGPT for video summaries (p2186925439 0:15-0:16)

Verbatim feedback prompt; apply manually (p2186925447 0:00-0:02)

Blog, newsletter, CTA, shorts prompts (p2186925451 0:00-0:04)

↓ Full write-up of this concept

07The 5-part ad formula, three enhancers, and three AI strategiesHook > Problem/Desire > Solution > Offer > CTA + proof/urgency/guarantee;

Hook > Problem/Desire > Solution > Offer > CTA + proof/urgency/guarantee; strategies: worksheet+prompt (Project), your hook + AI body, remix a winner + new hooks.

Five parts and why each (p2190229233 0:00-0:03)

Hook swap turned a loser into the top ad (p2190229233 0:01)

Enhancers: social proof, urgency, guarantees (p2190229233 0:05-0:07)

Verbatim copywriter prompt; Project as ad machine (p2190229233 0:08-0:10)

Foundation hook example ($6 -> $50k) (p2190229233 0:10-0:11)

Verbatim remix prompt; hook variations (p2190229233 0:11-0:12)

↓ Full write-up of this concept

The concepts in full

01

Thumbnail theory: stop the scroll, busy is bad, four words max, engineer curiosity

A fishing lure alone on a table stops nobody. A person holding the fish does - and then the lure becomes the question.

Rule 1, stop the scroll: 'If your audience can't figure out what they're looking at... you've already lost them' - anchor unfamiliar things with a relatable element. Rule 2, busy is bad: one primary subject (usually a person - people respond to faces and emotion), at most one secondary object that 'connects the dots', a palette of one or two complementary colours. Rule 3, text: none ideally, four words or fewer if needed - no-text thumbnails cost more storytelling (staged photos, Photoshop). Rule 4, curiosity: a confused or emotional face (beware fatigue), two objects that do not belong together, arrows or circles to a detail, a short provocative claim ('Big Mistake', 'Yikes'). Evidence: a simplified microphone thumbnail took 44.6% watch share vs 29%; 'Yikes' beat the same image without it; a student's NFL thumbnail went from 0.5% to 4.1% CTR after simplifying; a podcast thumbnail with illegible guest names critiqued against a recognizable claim.

Why it matters

Paul's client thumbnails and ad stills are judged by these same four rules; the split-test numbers make the case to a skeptical client.

02

Thumbnails in Nano Banana Pro, and the Triple Threat GPT

how-to

Do not chase the perfect prompt. Get a base image, then tell it to move you closer and enlarge the icons.

One high-quality reference (a character sheet if likeness is weak - see the clone lesson); a plain prompt - verbatim example: 'create a YouTube thumbnail featuring the man from the reference image. Background is a gradient blue... The man is on the left... expression slightly confused with an eyebrow raised... On the right side there are logos for CapCut, DaVinci Resolve and Clipchamp. In the top right... large easy to read text that says which is best... modern YouTube style... not too busy'; 16:9, 2K, two generations; iterate by downloading, re-uploading as the reference and giving edit-only instructions (move closer, enlarge, add stroke, change font). Multiple references combine a person with a real product (Sony ZV-E10 Mark II). The Triple Threat Thumbnail custom GPT: from scratch (one sentence in, three Nano Banana prompts out), improve your existing idea's prompt, or inspiration from another creator's thumbnail (download via youtube-thumbnail-grabber.com, upload, get a rewritten prompt, generate something new - never paste your face onto Ryan Trahan's design).

Do it in this order
Why it matters

Ten minutes per thumbnail, with an ethics line Paul can state to clients.

03

Train AI to write like you: two prompts and a correction habit

Paste one script you actually wrote. From then on it writes like you, formatted for a teleprompter.

Prompt 1 with a self-written script attached: 'Can you analyze my writing style? Add it to your memory and in the future write content similar to how I wrote this.' Prompt 2: 'moving forward, anytime I ask you to write a script, please format it in a teleprompter ready style so I can read it word for word to a camera.' Then correct as you go ('epic', not 'incredible') and tell it to remember; feed more documents over time. No sample? Write a short original piece or dictate a story via phone voice-to-text.

Why it matters

Generic AI copy is the most common client complaint; this is the two-minute fix.

04

Ideas -> titles -> outline -> script, one section at a time

how-to

'ChatGPT is like a muscle. The more we ask of it in one single request, the more it's going to start to fatigue.' So ask for the hook. Then the next section.

Seed a fresh thread with channel description, audience, values and the 'channels your audience watches' from YouTube Studio Analytics > Audience (or aspirational channels if you have none). 'Can you help me come up with 10 video ideas each with their own unique, engaging and curiosity driven title?' - then ten more on a subtopic. Pick one: 'Can you please come up with five variations of this title? Each one engaging, curiosity driven and under 60 characters' (so it is not truncated); ask which is most optimized for YouTube search. Outline it: 'outline a YouTube script for a video titled 5 Mistakes Beginner YouTubers Typically Make'; edit the outline by hand. Draft: 'Write a teleprompter ready YouTube script in my personal writing style based on this outline... under 10 minutes' for short pieces; for anything long, section by section - 'Write just the hook in intro for this video in my personal style teleprompter ready format' - each prompt told to build on the previous section. Read it aloud to catch stumbles and build muscle memory. Generalizes to shorts, email and blog.

Do it in this order
Why it matters

A repeatable production line Paul can hand to a client's marketing person verbatim.

05

ChatGPT Canvas as the writing desk (create it directly, or the tools vanish)

Open a reply into Canvas with the button and you get a text box. Create the Canvas first and you get a rewrite toolkit.

Canvas is 'an interactive workspace inside ChatGPT designed specifically for long form content' - a persistent document rather than disappearing replies. Tools: highlight-and-refine, rewrite for tone (professional/casual/humorous), expand with an example, adjust length, change reading level, Add Final Polish, add or remove emojis. His workflow: write the outline by hand, open Canvas (button or /canvas), 'please make a teleprompter ready script in my personal writing style following the outline attached below. The title of the video for context is this.' Caveat: converting a normal reply into Canvas via its button gives manual editing only; the AI tools need Canvas created directly.

Why it matters

The one UI gotcha that makes people think Canvas does nothing.

06

Research with seeded memory, feedback without surrender, and the repurposing cascade

'If chat doesn't know the answer, it will tell you' - as long as you asked it to be honest about statistics.

Research: seed memory with channel description, similar channels and top videos; 'what are trending topics in my niche' (he prefers evergreen); constrained picks ('the five best cameras from low to high price' with bands); comparison tables with added columns; dual-perspective paragraphs to counter bias; 'explain video frame rate to me as if I were a beginner. Include a visual analogy'; statistics with honest 'no data' answers; article summaries as tables; FAQs as a completeness checklist. Gap: ChatGPT cannot read a YouTube URL - NoteGPT (15 free/month, mind map). Feedback: 'can you read through my YouTube video script and provide feedback, check for any factual inaccuracies and also provide feedback on the writing style and any other areas that could use improvement' - strengths, visual suggestions, under-explained terms ('Open Gate'), sections to condense, CTA reminders; he applies by hand - 'too much of a control freak.' Repurpose: 'analyze this YouTube script and repurpose it into a blog post' (Google traffic; the embedded video 'YouTube also really likes'), then 'for an email newsletter... much shorter', add 'a call to action... to watch the full YouTube video', then 'three to five engaging short form video ideas' with hook and CTA.

Why it matters

One recording becomes a blog, a newsletter and five shorts - the content multiplier every small client needs.

07

The 5-part ad formula, three enhancers, and three AI strategies

how-to

Swap one hook and the worst ad in the account became the best. Everything after the first sentence is contingent on it.

Hook (first sentence or two), Problem or Desire ('people only trust solutions from other people who understand their problems'; pain moves people more than desire), Solution in the customer's language, Offer that stacks value into a no-brainer ('it costs less than a dinner date with your grandma'), explicit CTA. Enhancers: social proof (specific results, or the founder's story), urgency (deadlines or 'don't fall behind'), guarantees (refund or results-based). Strategy 1, full AI: fill the worksheet, then 'act as a professional ad copywriter. Based on the details below, write a high converting Facebook ad in my personal writing style using the Hook Problem Solution, Offer details and Call to Action framework. Make it emotionally engaging in under 150 words' (change the count); save worksheet + prompt as a ChatGPT Project - an 'ad writing machine.' Strategy 2, his most used: give your own hook ('My first AI ad cost $6 to make, but went on to generate over $50,000 for my business in the first month') and let AI finish under a cap. Strategy 3: 'below I pasted the script for an ad that was very successful. Please analyze this script and write a new and unique ad that builds off the same concept as the previous winner', then 3-4 hooks 'as shocking or as interesting as possible.'

Do it in this order
Why it matters

Pairs with Section 3's image ads: this is the copy, that is the picture, and the retainer is the business.

Tools referenced

ToolCoverageMomentContext
Nano Banana ProdemonstratedThumbnail generation and iteration
PromptEditdemonstratedFront end for Nano Banana Pro
ChatGPTdemonstratedMemory, Projects, Canvas; every writing workflow
YouTube StudiodemonstratedAnalytics > Audience > channels your audience watches
NoteGPTdemonstratedYouTube video summaries, 15 free/month
ClipchampmentionedLogo in the example thumbnail

Session materials

Archived locally on V: — click to open. Companion pages link to the LMS.

Action items

    Resources mentioned

    Resources
    • docVerbatim prompt bank (Sections 7-8)
    • docCourse downloads referenced

    Extraction notes

    This page was built from an auto-generated transcript, which garbles product and people's names. Those were corrected silently in everything above and logged here for transparency. The warnings flag claims that were true on the recording day but change fast.

    Transcript corrections applied

    The transcript saysThe trainer actually means
    note gtpNoteGPT
    Clip ChampClipchamp
    Sony ZV E10 Mark 2Sony ZV-E10 Mark II
    State of PAstate of pain
    nanobananaNano Banana Pro

    True on recording day — verify before relying