We analyzed 100
AI tutorial Shorts.
How do creators hook viewers?

01 · OPENING PATTERNS

The opening patterns in this 100-video sample

A Hook is the opening device that gives a viewer a reason to keep watching. We classified each video by its main opening device, using these definitions. When categories overlapped, the purpose took priority: a question setting up a test counted as an experiment; a question about a viewer's difficulty counted as pain led.

Valid n=100; missing primary classifications=0. No video received Counterintuitive as its primary label.

The labels describe the opening, rather than its performance. A delayed answer alone did not qualify a video as curiosity led.

02 · TIMING ACROSS 100 VIDEOS

When value becomes clear—and when useful content begins

We coded two moments across all 100 videos. Value clarity is when a viewer can identify the problem addressed or result offered. First substantive delivery is the first executable step, concrete answer, or meaningful task demonstration. A generic teaser does not count. A specific comparison result or a demonstrated input-to-output change can count without a click-by-click tutorial.

Each video has an estimated time range based on caption and video evidence. Wider ranges reflect less precise evidence; a range starting at zero means the event may be present in the opening.

The median value-clarity time falls within 0–6 seconds; the median first-delivery time falls within 7.5–13 seconds. We calculated these bounds from the medians of the 100 lower and upper endpoints. They describe uncertainty about when events occur, rather than recommended deadlines.

Valid n=100 for each metric; missing ranges=0. Exact onset times were not measured.

By ten seconds, 82 value-clarity ranges and 29 first-delivery ranges have ended. Another 18 value ranges and 41 delivery ranges span the ten-second mark: we cannot confidently place those events before or after it. Thirty delivery ranges begin after ten seconds.

The distinction appears in concrete cases. Jeff Su introduces the AI-memory outcome in an estimated 0–4-second window and starts the built-in-memory instruction in a 3–8-second window. Andy Stapleton's academic-prompting clip communicates its topic early, but the first document-placement instruction falls in a 27–33-second window.

Compare the two moments in the sample library, or download the complete timing table, including event descriptions and evidence references.

03 · NINE CASE STUDIES

Nine real openings worth studying

Kevin Stratvert's slide-editing tutorial pairs a clear on-screen headline with a direct capability promise. The task is concrete: create a presentation and change it by drawing on the slides.

The body enters Genspark, requests a six-slide cookie-baking presentation, shortens a subtitle, and uses an arrow to move an image. Those actions connect the opening to something the viewer can inspect. The reusable structure is one outcome, one tool, and a visible revision—not an abstract claim that AI saves time.

Opening promise · 0sThe workflow · 8.1s

Aurelius Tjin's Gamma presentation tutorial begins with a question about making presentations manually. Its opening headline already names the desired result.

The clip then moves to the tool, a prompt, a generated presentation, and simple edits. The referral offer is part of the content and is recorded in the research table. What transfers to another tutorial is the sequence: recognizable work, specific alternative, first useful step, inspectable result.

The result promise · 0s

Jeff Su's AI-memory tutorial organizes the body around three layers: built-in memory, projects, and a custom memory system. It provides an intelligible progression rather than a loose list of tips.

We marked delivery as partial because the opening's perfect-memory and never-correct-again scope exceeds what three examples establish. The useful writing lesson is to keep the progression and narrow the promise to what the video can demonstrate. A three-step structure does not require an unlimited guarantee.

The first step · 8.9s

Curious Refuge's AI-versus-human sound-design test asks a question that the body can actually address. It uses the same scene, plays the AI version and the human version, then gives a judgment.

We coded it as a challenge rather than a generic question. The comparison task supplies the structure. Its judgment is still the creator's assessment of these examples; it does not prove that one method always produces better sound design.

In Which AI is smartest?, Mrwhosetheboss uses a jar of jelly beans to make a model-comparison task visible. Each model is asked for an estimate; the final count supplies ground truth.

The ending answers the test that the opening establishes. It does not establish a universal smartest model. When adapting this structure, decide the input, scoring rule and ending before recording. Otherwise, a comparison can become a sequence of outputs without a conclusion.

The shared test input · 4s

Questions that AI can't answer begins with an image request involving a left-handed writer, then points out the mismatch. Egg counts and mirror reflections extend the pattern.

This is result-first structure built around a failure. A concrete prompt and an inspectable error give the explanation something to address. Our coding concerns presentation and payoff; it does not validate the video's simplified account of model internals.

Nate Herk's Claude-editing demonstration describes an iterative process: unclear captions become a contrast rule, static material becomes a motion requirement, and sync problems become checks.

The body gives each improvement a cause inside the workflow: a specific note became a persistent editing rule. That makes a repeatable structure for tutorials about refinement. Show the initial problem, state the change, and show what the change does. We cross-checked this source after detecting an inconsistent hosted keyframe; the disputed frame was not used as timing or visual evidence.

Design with Canva's photo-restyling walkthrough has no spoken opening sentence. A headline and a photo transformation introduce the task, then the screen recording uses arrows, prompts and interface actions to guide the viewer.

The ending shows the restyled composition in a before-and-after social mockup. This structure does not need a voiceover promise to make sense: the headline names the change, the screen supplies the procedure, and the ending returns to the change. The missing spoken sentence is recorded as absent, rather than reconstructed from music transcription.

The result · 81.2s

Andy Stapleton's Perplexity introduction workflow begins with the problem of not having enough references. It moves into a specific research prompt, shows the resulting draft and references, and then explains how he would use the references in a citation workflow.

He explicitly treats the output as a starting point rather than text to copy directly. The transferable structure is a concrete gap, a tool operation, an inspectable output and a bounded next step. Our study does not independently validate the generated references or treat the draft as finished academic work.

04 · SIX SCRIPT STRUCTURES

Six reusable script structures

Adapt these six templates from the observed structures with your own task, footage and proof.

1. 01

   One outcome, one tool

   YOUR OPENING

   Here is how to make [specific output] and change [one detail] in [tool].

   WHAT TO DELIVER

   The input, output and actual edit.

   Study the original example ↗
2. 02

   Pain to first useful click

   YOUR OPENING

   Still doing [recognizable task] manually? Here is the workflow I use.

   WHAT TO DELIVER

   A concrete path into the tool, not only a recommendation.

   Study the original example ↗
3. 03

   A numbered progression

   YOUR OPENING

   Use these three steps to manage [bounded problem].

   WHAT TO DELIVER

   Three distinct steps and an ending that matches the scope.

   Study the original example ↗
4. 04

   A bounded comparison

   YOUR OPENING

   Can [tools or methods] solve [the same task]? Here is how I will check.

   WHAT TO DELIVER

   Consistent inputs, inspectable outputs and a stated comparison rule.

   Study the original example ↗
5. 05

   Failure before explanation

   YOUR OPENING

   I asked for [specific output]. Here is the mismatch.

   WHAT TO DELIVER

   The input, failed output and an explanation within the evidence.

   Study the original example ↗
6. 06

   Revision loop

   YOUR OPENING

   My first attempt had [visible problem]. Here is the rule that changed it.

   WHAT TO DELIVER

   A before state, the change and a result the viewer can assess.

   Study the original example ↗

05 · PROMISE DELIVERY

Did the body fulfill the opening promise?

Our coding marked 88 videos as delivering their opening promise and 12 as partially delivering it. None was coded No or Unclear. Valid n=100; missing judgments=0.

This is an assessment of the relationship between opening and body. It does not certify product claims, audience satisfaction, or effectiveness. Partial often reflected a promise broader than the example could establish: perfect memory, universal superiority, total replacement, or an outcome deferred beyond the clip.

During scripting, write the opening and ending next to each other. A tutorial promising one transformation should show that transformation. A test needs an answer under its stated conditions. A broad claim needs either more evidence or narrower wording.

Two research tutorials show how scope can drift. One introduces five prompting improvements but supplies only the first in the clip. Another opens with a twelve-word prompt claim and ends after source-upload instructions, before supplying that prompt. Both are marked Partial.

06 · APPLY THE FINDINGS

Use the findings in your next tutorial

Choose the task before choosing the wording. Identify the input, output or failure that makes the task concrete. Then decide what the spoken line, first image and on-screen words each contribute.

Place the first executable instruction or meaningful demonstration on the timeline. Check whether the setup before it is necessary. Keep the ending visible while you edit the opening, so a more exciting line does not quietly promise a different video.

Use short-form Hook prompts to explore wording once the evidence is clear. After publishing, use your own analytics to test the opening.

07 · METHOD & SAMPLE

How we built the sample

We screened 60 channel records; 54 met the threshold of at least 100,000 subscribers. The final sample contains 100 videos from 30 channels, with one to five videos per channel.

Videos had to be published between October 27, 2025 and October 26, 2026 (Asia/Shanghai dates), run 15–90 seconds, and teach or demonstrate an identifiable AI-tool task through English spoken or written instruction. We excluded pure news, general theory, non-AI tutorials and promotions without a concrete demonstration. Sponsorship alone did not exclude a useful tutorial.

Across eligible accounts, 138 upload-list pages yielded 6,872 unique records. Of these, 5,910 fell inside the date window and 2,009 also met the duration rule. Metadata screening produced 510 candidates; 214 received content-screening decisions. The remaining candidates were not all reviewed.

We followed the recorded account-pool order and used the first 30 accounts with identified, accessible eligible material. We filled the sample in account rounds, taking the next newest identified eligible video from each account until reaching 100. The ceiling was ten per account; the actual maximum was five. There was no minimum view count or view-based ranking. Upload-page coverage was bounded, so this is not an exhaustive catalogue of eligible videos.

Public metadata was collected on October 27 and 29, 2026. Three included videos had at least one million public views; all three came from Mrwhosetheboss. We did not use this small subgroup to compare Hook performance.

An AI assistant assigned one primary Hook label, assessed promise delivery and estimated two event ranges using timestamped captions, whole-video analysis and selected extracted frames. Timing ranges were rounded outward to whole seconds. The median bounds use the medians of all lower and upper endpoints; they are not confidence intervals. Caption errors and interpretation can affect the labels and timings.

Download the 100-video sample data for video IDs, publication dates, view snapshots, labels and evidence notes. Each included video also has a source-linked breakdown in the sample library.

08 · LIMITS & EVIDENCE

What this research cannot establish

This is an exploratory sample of established channels, not a random or representative sample of creators. Videos from the same channel may share an opening style. These patterns cannot establish which Hook causes better performance.

All 100 videos received whole-video analysis. Ninety-seven have usable spoken captions; three use English screen instructions without voiceover. Detailed visual evidence covers a subset of the sample, and timing ranges remain estimates. No independent inter-rater agreement statistic was calculated.

Bounded upload-page coverage, metadata screening and access failures can affect which videos enter the sample. Source links, screening decisions and evidence notes are retained for checking the analysis.

Public views cannot reveal retention, distribution source or why a video succeeded. YouTube provides the audience-retention report in video-level Analytics. Use the observed structures as ideas to test against your own audience data.