Skip to content

mbfinotti/advertising-skills / skills/ad-hook-analyzer

ad-hook-analyzer

Verified

Score and force-rank the openings of candidate video ads - from a script, transcript, storyboard, shot list, or a description of a finished cut - to decide which hooks deserve test budget, returning a ranked shortlist within the batch rather than an absolute performance prediction. Use whenever the user mentions a hook, the first 3 seconds, an ad opening, hook rate, thumbstop, attention scoring, or asks which hook to test or whether a hook works - even if they never say 'hook analyzer'. Covers B2B and B2C on any platform. Ends at the ranking: writing the scripts themselves is mbfinotti/advertising-skills@ugc-ad-scripts and static copy is mbfinotti/advertising-skills@ad-copy-variants.

FileKindSize
SKILL.mdskill md18.9 KB
evals/evals.jsonconfig39.4 KB
references/examples.mdmarkdown12.1 KB
references/hook-patterns.mdmarkdown6.1 KB
references/platform-notes.mdmarkdown6.2 KB
references/scoring-rubric.mdmarkdown10.0 KB

references/scoring-rubric.md

View on GitHub
# Scoring Rubric - Seven Dimensions

Band every gate-passing candidate **strong / adequate / weak** on each dimension. **Never invent numeric sub-scores, percentages, or a composite total** - the bands exist precisely because a number would claim precision the evidence does not support, and because human hook-scoring has poor inter-rater reliability even among professionals. When two bands both seem defensible from the input available, band `adequate`, and flag the missing input (e.g. "storyboard doesn't show the first frame - a frame grab would resolve time-to-signal").

Bands feed the pairwise ranking as evidence, not arithmetic. Never rank by counting strongs.

**The seven dimensions are unranked on purpose** - no weights, no importance order, no fix-first
sequence, and nothing below should be read as one. With hook rate correlating **-0.19 with
ROAS**, no evidence supports claiming that any one dimension returns more performance per unit
of effort than another, so an ordering here would be invented precision dressed as guidance. The
ranking this skill does emit is between candidate openings, never between the axes used to judge
them; the ordered ladder in the skill's Raising a Rank section ranks _revisions_, whose effort
differs by orders of magnitude and is observable before launch.

## 1. Time-to-signal

**Measures**: how many seconds pass before the viewer knows what is in it for them - not when something _happens_, but when the personal relevance lands.

**Evidence**: feed attention decays steeply - roughly 80% of the audience still present in second 1, 50% by second 2, 20% by second 3 (Karen Nelson-Field / Amplified Intelligence, human-attention datasets across Facebook, YouTube, Instagram, TV). Memory reliably encodes only after roughly 2.5 seconds of active (eyes-on) attention; Amplified/VCCP (May 2025, 20,000+ views of 72 digital video ads) found 1.5 seconds can suffice when distinctive brand assets are deployed well. The practical read: the value signal must land while the majority is still there - by around second 2.

- **Strong**: the "what's in it for me" is legible in the first frame or first line - a viewer who saw only second 1-2 could state the promise.
- **Adequate**: the signal lands within the placement's hook window, but seconds 1-2 are spent on setup that earns nothing on its own.
- **Weak**: throat-clearing - greeting, logo card, scene-setting, or a tease with no defined shape - and the actual signal arrives after the window, when most of the audience has gone.

**Ambiguous case**: scripts often under-specify the first visual frame. If the verbal line signals fast but the visual is unknown, band `adequate` and request the first-frame plan.

## 2. Sound-off legibility

**Measures**: how well the opening delivers its meaning with audio muted - beyond the pass/fail gate, which only asks whether it survives at all.

**Evidence**: feed video autoplays muted on Meta and LinkedIn (platform behaviour, verified); Meta's own creative guidance recommends carrying the opening with text, graphics, and captions. Note on folklore: the famous "85% of Facebook video is watched without sound" is a 2016 publisher-reported snapshot (Digiday, citing LittleThings and Mic) that Facebook never confirmed - do not cite it as a current or platform-verified figure. The design constraint stands on muted autoplay alone.

- **Strong**: first frame plus on-screen text carry the full promise; audio only adds texture. A muted viewer loses nothing essential.
- **Adequate**: the gist survives muting, but a meaningful layer (the specific claim, the punchline) lives only in the audio.
- **Weak** (usually a gate FAIL instead): meaning depends on voiceover or sound design; muted, the opening is wallpaper.

**Ambiguous case**: transcripts contain no visual/text information at all. Band `adequate` at best and say the input cannot support `strong`.

## 3. Audience qualification / self-selection

**Measures**: how precisely the opening lets the _right_ viewer recognise "this is for me" - and, equally, lets the wrong viewer scroll past. The goal is not maximum attention; it is qualified attention.

**Evidence**: Barry Hott (growth/creative consultant, ~$1B managed spend): "Higher CTR or hook rate (thumbstop) doesn't mean an ad will perform better or worse... Clickbait can get tons of clicks without generating any sales." The same logic drives the documented high-hook-rate/mediocre-ROAS failure accounts.

- **Strong**: the opening names or unmistakably implies the buyer's role, situation, or problem, and the disqualified viewer (per the Interview) has no reason to stay.
- **Adequate**: relevant to the buyer but broad - the right viewer might self-select, plenty of wrong viewers will too.
- **Weak**: pure spectacle or universal curiosity - it stops everyone, which is qualification for no one - or so narrow/inside that even the buyer doesn't recognise themselves.

**Ambiguous case**: when the Interview's "what disqualifies a viewer" answer is thin, band `adequate` and push the question again - this dimension cannot be banded without it.

## 4. Specificity

**Measures**: whether the opening makes a concrete, checkable claim or shows a concrete situation, versus generic category language.

**Evidence**: convergent practitioner consensus rather than a single controlled study - specific numbers, named situations, and checkable claims recur across practitioner hook analyses as separating working openings from generic ones, and generic category openings recur in every anti-pattern list. Treat as well-supported practice, not measured law.

- **Strong**: a claim you could check or a situation you could film - a number, a named context, a visible demonstration ("this skillet went through 1,000 dishwasher cycles").
- **Adequate**: a real benefit stated abstractly - true, relevant, but interchangeable with a competitor saying it.
- **Weak**: category wallpaper - "revolutionize your workflow," "the smarter way to X" - language any brand in the category could run unchanged.

**Ambiguous case**: a bold specific claim that the product cannot back is not `strong` specificity - it is a promise-payoff gate problem. Check the gate before banding.

## 5. Brand legibility timing

**Measures**: whether the choice of when and how the brand becomes legible in the opening fits the stated objective and platform. This is a genuine, unresolved tension - band the _fit of the choice_, not adherence to a rule.

**Evidence, deliberately in tension**: opening on a logo card signals "this is an ad" at the worst moment (documented anti-pattern); yet withholding branding costs brand linkage - TikTok reports a 17-point drop in brand linkage when branding is left to the end (TikTok-cited figure, asserted rather than a controlled finding). Amplified/VCCP (2025) found distinctive brand assets, used well, let memory encode in as little as 1.5 seconds - which argues for early _distinctive-asset_ branding without a logo card. The evidence does not settle the question; objective and platform decide it.

- **Strong**: a deliberate choice matching the objective - direct-response cold traffic: brand cues woven in (product in hand, distinctive colour/asset) without an ad-announcing logo open; brand/memory objective (most B2B): distinctive assets legible early, per the linkage evidence.
- **Adequate**: branding present and findable, but by default rather than by design - neither hurting nor working.
- **Weak**: the accidental extremes - a logo card or brand announcement as the opening beat with a DR objective, or a brand-building objective where nothing identifies the brand until the end card.

**Ambiguous case**: if the user has not stated an objective, this dimension cannot be banded honestly - ask, don't assume DR.

## 6. Promise-payoff continuity

**Measures**: beyond the pass/fail gate - how tightly the opening's promise _is_ the ad's payoff and the offer's substance, versus merely not contradicting them.

**Evidence**: the documented failure this guards against - "I've run campaigns with 43% hook rates and mediocre ROAS. The hook stopped the scroll... but it didn't qualify the right viewer or set up the offer" (named practitioner account, Imagine.art). The same source describes teams lifting hook rate from 18% to 37% across 40 variations with no ROAS change.

- **Strong**: the opening is the offer in miniature - the promise made in the window is exactly what the ad demonstrates and the offer sells; a viewer who converts got what second 1 promised.
- **Adequate**: promise and payoff are related but the opening oversells or angles away - the ad delivers a cousin of the promise.
- **Weak** (usually a gate FAIL): bait - the promise exists to stop the scroll and the body pivots to something else.

**Ambiguous case**: when only the opening was supplied, the gate and this band cannot be judged - require at least a summary of the full ad and the offer before scoring the batch.

## 7. Placement-native fit

**Measures**: whether the opening is built for the stated placement's hook window and viewing context, or transplanted from another format.

**Evidence**: the windows differ materially (platform-defined, verified):

- Feed placements: ≈ first 3 seconds.
- YouTube skippable in-stream: ≈ first 5 seconds, before the skip button.
- Bumpers and non-skippable formats: forced views, so scroll-stopping logic does not apply at all. Only continuity, branding, and qualification carry meaning.

Sound norms, text-safe zones, and pacing also differ by placement.

- **Strong**: window, first-frame composition, text placement, and pacing all built for the stated placement.
- **Adequate**: transferable with light edits - right length, but text sits in another platform's safe zone or pacing follows another format's rhythm.
- **Weak**: a transplant - a 15-30 second cold-open structure dropped into a 3-second feed window, a watermarked cross-post, or a feed-style scroll-stopper submitted for a forced-view slot.

**Ambiguous case**: if the user plans multiple placements for one batch, score against the primary placement and note that the ranking holds only for it - a re-rank per placement is a separate pass.
finishing-a-development-branchobra/superpowersUse when implementation is complete, all tests pass, and you need to decide how to integrate the workStars: 322kUpdated Claude CodeCodexVerifiedInstall
systematic-debuggingobra/superpowersUse when encountering any bug, test failure, or unexpected behavior, before proposing fixesStars: 322kUpdated Claude CodeCodexVerifiedInstall
test-driven-developmentobra/superpowersUse when implementing any feature or bugfix, before writing implementation codeStars: 322kUpdated Claude CodeCodexVerifiedInstall
skill-creatoranthropics/skillsCreate new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy.Stars: 198kUpdated Claude CodeCodexVerifiedInstall