wizz-method 2.16.0 → 2.17.1

package/src/skills-lib/wizz-video-styles/references/styles/motion-editorial.md

@@ -0,0 +1,69 @@
+# Motion editorial (Greg Isenberg)
+
+Measured on 3 shorts (en) in: dev / AI / build in public (EN).
+
+## Measured
+
+| Metric | Value |
+|---|---|
+| Reach (views) | max 829,395 · median 642,459 |
+| Cuts per 10 s | min 1.2 · median 1.71 · max 1.95 |
+| Median scene length | 4.1 s |
+| Music bed (continuous / faint / none) | 3 / 0 / 0 |
+| Cuts with a sound effect on them | 3/23 (13%) |
+
+Cuts come from scene detection (it undercounts overlays and cards); music and SFX from an audio heuristic that classifies type, never names the sound ([../research-method.md](../research-method.md)).
+
+## Material it needs
+
+Only a voice track. No one on camera required (a short face insert is optional). App or product screens as screenshots.
+
+## When it fits
+
+Explaining a system, a workflow, a tool or a framework; tech, SaaS, AI, productivity, finance apps, B2B. Calm authority, saves and comments.
+
+## Recipe
+
+**Look**
+
+- Light warm background (cream, sand, light grey) with a faint square grid and a soft vignette.
+- One object at a time in the centre: dark rounded cards with deep soft shadows, pill-shaped inputs, file chips, a phone card holding a real screenshot.
+- An accent shape in one warm colour (orange, copper) that recurs: rays, a ring, a starburst.
+- Dashed connectors drawing between chips; text typed inside pills.
+
+**Palette:** Background #EEE7DA, ink #1A1A1A, card #1C1916 → #2A2521, card text #DFD8C6, one accent (#B5541C / #C9622A / #E39A6C). Swap the accent for the brand colour; keep the light background.
+
+**Typography:** Serif italic display for headlines and pill text (Playfair Display italic 600/700); sans for small UI; mono for code.
+
+**Motion**
+
+- Each scene enters with a spring (scale 0.94 → 1, rise 70 px) and leaves rising and fading; scenes overlap ~0.1 s, so there is no hard cut.
+- Slow camera push of 3 to 4 % across a scene.
+- Objects transform into the next object (pill becomes card, card becomes phone) instead of cutting.
+- Counters tick up to real numbers; typing in pills; connectors draw in.
+
+**Captions:** Small or none. If used: one line of 2 to 4 words, sans bold ~50 px, ink colour on the light background, active word in the accent. Headline in serif italic, two-tone (ink + accent), only at key moments.
+
+**Hook (0 to 3 s):** A serif headline of the claim with one object animating in under it (0 to 3 s). A real number works well.
+
+**CTA:** The CTA word typed in a pill with the accent shape lighting up; one line of serif below.
+
+**Music:** Continuous bed, calm (piano, ambient, soft electronic), around 0.2 of the voice level with fade in/out.
+
+**SFX:** Rare: a soft whoosh on the 2 or 3 biggest transformations, at most ~1 per 10 s. Light typing under one typed pill. Nothing on every scene.
+
+## Do not
+
+- Full-screen browser screen recordings (they can come out stretched); use real screenshots inside a phone card.
+- Giant outlined captions; they fight the editorial look.
+- A sound on every scene change; pops, dings and impacts in sequence.
+
+## Watch before editing
+
+- [Greg Isenberg](https://www.youtube.com/shorts/ovLAIhbk3ek) · 829,395 views
+- [Greg Isenberg](https://www.youtube.com/shorts/yJK5GueSHmU) · 642,459 views
+- [Greg Isenberg](https://www.youtube.com/shorts/5G9PAIUs5Pk) · 458,691 views
+
+## Template
+
+`assets/remotion/editorial/` (Remotion). Copy it into the project's `src/`, set the theme and the scenes; see its README.

package/src/skills-lib/wizz-video-styles/references/styles/podcast-clip.md

@@ -0,0 +1,61 @@
+# Podcast or interview clip
+
+Measured on 12 shorts (en, pt) in: dev / AI / build in public (EN), personal finance / investing, product / SaaS / marketing / business / law, health / fitness / nutrition.
+
+## Measured
+
+| Metric | Value |
+|---|---|
+| Reach (views) | max 9,846,107 · median 216,533 |
+| Cuts per 10 s | min 0.0 · median 1.91 · max 4.42 |
+| Median scene length | 3.5 s |
+| Music bed (continuous / faint / none) | 0 / 6 / 6 |
+| Cuts with a sound effect on them | 7/124 (6%) |
+
+Cuts come from scene detection (it undercounts overlays and cards); music and SFX from an audio heuristic that classifies type, never names the sound ([../research-method.md](../research-method.md)).
+
+## Material it needs
+
+A recorded conversation or talk with at least one good camera; a second angle helps.
+
+## When it fits
+
+Finance, health, business, mindset; any expert talking. Reuses long-form content.
+
+## Recipe
+
+**Look**
+
+- Two cameras alternating speakers in a dark studio; channel badge in a corner.
+- Variants: letterbox with a title bar on top; tier list or ranking being filled over the clip; reaction picture-in-picture; black background with the clip in a window and tools stacking (Starter Story).
+
+**Palette:** Studio darks; captions white with yellow/red keywords; title bar white with black text.
+
+**Typography:** Condensed bold uppercase captions; plain bold sans titles.
+
+**Motion**
+
+- Cuts follow the speaker; inserts of the thing being discussed (product, card, chart) full frame for 1 to 2 s.
+
+**Captions:** Word by word, large, uppercase, white with outline, keyword in yellow or red, lower-centre.
+
+**Hook (0 to 3 s):** Start mid-sentence on the strongest line, or with the title/ranking on screen.
+
+**CTA:** Spoken or a title card; often none (the clip ends on the punchline).
+
+**Music:** None or faint.
+
+**SFX:** Almost none on cuts.
+
+## Do not
+
+- Starting with greetings or context.
+- Clips that need the long episode to make sense.
+
+## Watch before editing
+
+- [Doctor Mike](https://www.youtube.com/shorts/cnyqT48dyjE) · 9,846,107 views
+- [Graham Stephan](https://www.youtube.com/shorts/K7UvJQsjMOM) · 1,238,146 views
+- [Theo - t3․gg](https://www.youtube.com/shorts/vgdpLKxS8dY) · 476,638 views
+- [Graham Stephan](https://www.youtube.com/shorts/V1hW3DH69pw) · 371,080 views
+- [Alex Hormozi](https://www.youtube.com/shorts/57_uoyn15_w) · 332,850 views

package/src/skills-lib/wizz-video-styles/references/styles/pov-vlog.md

@@ -0,0 +1,57 @@
+# POV vlog (Drew Binsky)
+
+Measured on 3 shorts (en, pt) in: food / recipes / lifestyle / travel, beauty / fashion / local services (clinic, salon, architecture, real estate).
+
+## Measured
+
+| Metric | Value |
+|---|---|
+| Reach (views) | max 8,726,757 · median 1,394,792 |
+| Cuts per 10 s | min 0.39 · median 2.71 · max 6.38 |
+| Median scene length | 3.27 s |
+| Music bed (continuous / faint / none) | 2 / 1 / 0 |
+| Cuts with a sound effect on them | 2/56 (4%) |
+
+Cuts come from scene detection (it undercounts overlays and cards); music and SFX from an audio heuristic that classifies type, never names the sound ([../research-method.md](../research-method.md)).
+
+## Material it needs
+
+Handheld footage of a place or experience, wide lens, the creator's voice.
+
+## When it fits
+
+Travel, experiences, behind the scenes, 'a day at'.
+
+## Recipe
+
+**Look**
+
+- Wide-angle phone footage, quick cuts, a selfie at the end; emphasis word in red.
+
+**Palette:** The place; captions white with shadow; red emphasis.
+
+**Typography:** Bold sans with black outline.
+
+**Motion**
+
+- Quick cuts (median ~2.7 per 10 s, up to ~6.4); question-mark stickers; movement in every shot.
+
+**Captions:** 1 to 3 words, white bold, lower-centre, emphasis in red.
+
+**Hook (0 to 3 s):** The most striking shot of the place plus a curiosity title in 0 to 1 s.
+
+**CTA:** Selfie to camera at the end.
+
+**Music:** Continuous or faint bed.
+
+**SFX:** Almost none on cuts.
+
+## Do not
+
+- Static shots; this style is movement.
+
+## Watch before editing
+
+- [Drew Binsky](https://www.youtube.com/shorts/d9w9tME0aEY) · 8,726,757 views
+- [Mari Maria](https://www.youtube.com/shorts/BuQbDL9F9eY) · 1,394,792 views
+- [Cata e Davi](https://www.youtube.com/shorts/kdC1VJtz0EY) · 624,507 views

package/src/skills-lib/wizz-video-styles/references/styles/quadro-branco.md

@@ -0,0 +1,60 @@
+# Whiteboard or hand drawing
+
+Measured on 5 shorts (en, pt) in: dev / AI / build in public (EN), dev / AI / tech business (BR), product / SaaS / marketing / business / law, health / fitness / nutrition.
+
+## Measured
+
+| Metric | Value |
+|---|---|
+| Reach (views) | max 4,438,595 · median 608,820 |
+| Cuts per 10 s | min 0.0 · median 1.31 · max 1.96 |
+| Median scene length | 6.12 s |
+| Music bed (continuous / faint / none) | 2 / 1 / 2 |
+| Cuts with a sound effect on them | 0/28 (0%) |
+
+Cuts come from scene detection (it undercounts overlays and cards); music and SFX from an audio heuristic that classifies type, never names the sound ([../research-method.md](../research-method.md)).
+
+## Material it needs
+
+A board, a notebook or a drawing app, and someone explaining. Face optional.
+
+## When it fits
+
+Concepts, frameworks, health explanations, programming ideas. Long single takes are acceptable here.
+
+## Recipe
+
+**Look**
+
+- Hand drawing in a notebook with highlighters (Filipe Deschamps), a board in one take (Dayan Siebra), white lines drawn on black (ThePrimeagen), a flip chart behind the speaker (Codie Sanchez).
+- A fixed question or title box at the top for the whole video.
+
+**Palette:** Paper white or board black; one neon highlighter; ink colours of the markers.
+
+**Typography:** Handwriting for the content; plain sans in a white box for the fixed title.
+
+**Motion**
+
+- Drawing appears progressively; few cuts; long takes (a median scene of several seconds).
+
+**Captions:** Often none beyond the drawing; when present, small and plain.
+
+**Hook (0 to 3 s):** The provocative question fixed on screen from frame one.
+
+**CTA:** Spoken; sometimes the drawing ends on the CTA.
+
+**Music:** Mixed; often none.
+
+**SFX:** None measured on cuts.
+
+## Do not
+
+- Drawing that the viewer cannot read on a phone (write big).
+
+## Watch before editing
+
+- [ThePrimeagen](https://www.youtube.com/shorts/_ebVsYQou8A) · 4,438,595 views
+- [Filipe Deschamps](https://www.youtube.com/shorts/wydSHobAaZQ) · 1,142,498 views
+- [Dayan Siebra](https://www.youtube.com/shorts/V4v1hg69XRI) · 608,820 views
+- [Filipe Deschamps](https://www.youtube.com/shorts/MLDH_t3Rpms) · 335,545 views
+- [Codie Sanchez](https://www.youtube.com/shorts/cLfZgtGNNx4) · 145,271 views

package/src/skills-lib/wizz-video-styles/references/styles/talking-head-cards.md

@@ -0,0 +1,62 @@
+# Talking head with cards and b-roll
+
+Measured on 13 shorts (en, pt) in: dev / AI / build in public (EN), dev / AI / tech business (BR), personal finance / investing, product / SaaS / marketing / business / law, health / fitness / nutrition, beauty / fashion / local services (clinic, salon, architecture, real estate), education / science / comedy.
+
+## Measured
+
+| Metric | Value |
+|---|---|
+| Reach (views) | max 2,181,318 · median 431,649 |
+| Cuts per 10 s | min 0.0 · median 0.45 · max 5.31 |
+| Median scene length | 9.47 s |
+| Music bed (continuous / faint / none) | 5 / 6 / 2 |
+| Cuts with a sound effect on them | 20/174 (11%) |
+
+Cuts come from scene detection (it undercounts overlays and cards); music and SFX from an audio heuristic that classifies type, never names the sound ([../research-method.md](../research-method.md)).
+
+## Material it needs
+
+A person on camera plus things to show: news headlines, article screenshots, charts, product shots, stock b-roll.
+
+## When it fits
+
+News and commentary (finance, tech, science), explainers with evidence, any niche that needs proof on screen.
+
+## Recipe
+
+**Look**
+
+- Face in selfie or studio; cards slide over the frame: headline screenshots with highlighter, numbered section titles, charts, stickers.
+- Variants: split screen (archive images on top, presenter below), face in a bubble over full-screen b-roll.
+
+**Palette:** Neutral room; white cards; highlighter yellow; one alert colour (red) for numbers and arrows.
+
+**Typography:** Condensed bold uppercase for section titles; medium sans for captions (yellow with outline on BR channels).
+
+**Motion**
+
+- Cards enter over the face rather than cutting away; cuts mostly when the topic changes.
+- Face bubble stays on from after the hook to the CTA when b-roll is full screen.
+
+**Captions:** Lower-centre, yellow or white with keyword emphasis; section number big at the top when the piece is a list.
+
+**Hook (0 to 3 s):** A thumbnail-like card in the first frames ('THIS BANK WILL BE NEXT?') with a red arrow or the key number.
+
+**CTA:** Back to the face full screen, no card on top.
+
+**Music:** Faint bed in most; continuous in the faster ones.
+
+**SFX:** A soft whoosh or paper sound when a card enters, on a minority of cards.
+
+## Do not
+
+- Cards covering the face for long stretches.
+- Unreadable article screenshots (crop to the line that matters and highlight it).
+
+## Watch before editing
+
+- [Manual do Mundo](https://www.youtube.com/shorts/q1DLHGpLSss) · 2,181,318 views
+- [Katia Haranaka](https://www.youtube.com/shorts/1NEVJoOAzcE) · 1,960,278 views
+- [Bruno Perini - Você MAIS Rico](https://www.youtube.com/shorts/xZYgNVefQsQ) · 1,159,863 views
+- [NetworkChuck](https://www.youtube.com/shorts/lO961HRQn5Q) · 815,553 views
+- [Bruno Perini - Você MAIS Rico](https://www.youtube.com/shorts/s5cwz0SeU5I) · 720,508 views

package/src/skills-lib/wizz-video-styles/references/styles/talking-head-legenda.md

@@ -0,0 +1,62 @@
+# Talking head with highlighted captions
+
+Measured on 21 shorts (en, pt) in: dev / AI / build in public (EN), dev / AI / tech business (BR), personal finance / investing, product / SaaS / marketing / business / law, beauty / fashion / local services (clinic, salon, architecture, real estate), education / science / comedy.
+
+## Measured
+
+| Metric | Value |
+|---|---|
+| Reach (views) | max 9,809,160 · median 382,506 |
+| Cuts per 10 s | min 0.0 · median 0.92 · max 5.69 |
+| Median scene length | 7.42 s |
+| Music bed (continuous / faint / none) | 0 / 6 / 15 |
+| Cuts with a sound effect on them | 17/188 (9%) |
+
+Cuts come from scene detection (it undercounts overlays and cards); music and SFX from an audio heuristic that classifies type, never names the sound ([../research-method.md](../research-method.md)).
+
+## Material it needs
+
+A person on camera talking (selfie, studio, stream, stage). Nothing else.
+
+## When it fits
+
+Opinion, news, advice, reaction, review, beauty talk; every niche in the set has it. Reach comes from the content, not the edit.
+
+## Recipe
+
+**Look**
+
+- One framing (selfie or studio), sometimes a second angle; jump cuts remove pauses.
+- A headline box (white or black) in the first seconds; captions word by word.
+
+**Palette:** Whatever the room is; captions white with outline, one accent for the keyword (yellow is the BR default; red, green or purple elsewhere).
+
+**Typography:** Bold sans uppercase for captions; condensed bold for the headline box; some finance/BR channels use a white serif.
+
+**Motion**
+
+- Few cuts (median ~1 per 10 s); punch-in zooms on emphasis; occasional reaction insert.
+
+**Captions:** Word by word or 1 to 3 words, bold, lower-centre (above the platform UI), keyword in the accent colour.
+
+**Hook (0 to 3 s):** The headline box with the claim or question in the first frame, spoken at the same time.
+
+**CTA:** Spoken to camera; caption repeats the keyword.
+
+**Music:** None or faint in most; the voice carries it.
+
+**SFX:** Rare; an occasional pop on a caption keyword.
+
+## Do not
+
+- Leaving pauses and 'uhs' in.
+- Captions under the like/comment buttons.
+- Music that competes with the voice.
+
+## Watch before editing
+
+- [Thiago Concer](https://www.youtube.com/shorts/-xP6HufvUOk) · 9,809,160 views
+- [ThePrimeagen](https://www.youtube.com/shorts/M-VU0fLjIUU) · 5,401,385 views
+- [rodrigo vizu](https://www.youtube.com/shorts/zcXXksbFwIQ) · 3,759,620 views
+- [Felipe Neto](https://www.youtube.com/shorts/7U4hRvg0vXw) · 1,660,932 views
+- [Mari Maria](https://www.youtube.com/shorts/rNJNJBpz1P8) · 1,419,599 views

package/src/skills-lib/wizz-video-styles/references/styles/tour.md

@@ -0,0 +1,59 @@
+# Walkthrough or tour (Enes Yilmazer, Ryan Serhant)
+
+Measured on 3 shorts (en, pt) in: food / recipes / lifestyle / travel, beauty / fashion / local services (clinic, salon, architecture, real estate).
+
+## Measured
+
+| Metric | Value |
+|---|---|
+| Reach (views) | max 1,051,145 · median 370,693 |
+| Cuts per 10 s | min 0.4 · median 2.88 · max 4.47 |
+| Median scene length | 2.1 s |
+| Music bed (continuous / faint / none) | 0 / 1 / 2 |
+| Cuts with a sound effect on them | 8/90 (9%) |
+
+Cuts come from scene detection (it undercounts overlays and cards); music and SFX from an audio heuristic that classifies type, never names the sound ([../research-method.md](../research-method.md)).
+
+## Material it needs
+
+A place to walk through (property, studio, store) and someone guiding.
+
+## When it fits
+
+Real estate, architecture, travel destinations, venues, clinics, stores; local services.
+
+## Recipe
+
+**Look**
+
+- Walking through the space facing the camera; a seated close; an insert of a phone/app when relevant; or one long take of the agent talking with a title box.
+
+**Palette:** The space; captions white; brand red for logos.
+
+**Typography:** Plain or condensed sans; title in a white box at the start.
+
+**Motion**
+
+- Walking tours cut every ~2 s (median scene 2.1 s) between rooms or spots; the single-take agent talk is the slow variant.
+- The walk itself is the motion; no graphic transitions.
+
+**Captions:** 1 or 2 words, white, centre or middle third.
+
+**Hook (0 to 3 s):** The most impressive room first, or the quote of the claim in a title box.
+
+**CTA:** Spoken; contact on screen.
+
+**Music:** None measured.
+
+**SFX:** None measured on cuts.
+
+## Do not
+
+- Graphic transitions between rooms; cut on the walk.
+- Measured reach is modest (hundreds of thousands to ~1M), not tens of millions.
+
+## Watch before editing
+
+- [Viaje Por Conta](https://www.youtube.com/shorts/3mmhzj4y_78) · 1,051,145 views
+- [Enes Yilmazer](https://www.youtube.com/shorts/FQYtkO9SpmU) · 370,693 views
+- [Ryan Serhant](https://www.youtube.com/shorts/fX7uB6IobEk) · 150,846 views

package/src/skills-lib/wizz-video-styles/scripts/analyze.sh

@@ -0,0 +1,66 @@
+#!/usr/bin/env bash
+# Measures one public YouTube short: download (720p), cuts (scene detection),
+# audio (music bed, SFX type heuristic, loudness) and a 16-frame contact sheet.
+# Usage: analyze.sh <youtube_id> <out_dir>  →  <out_dir>/<id>.json + <id>-sheet.png
+# The .mp4 is deleted at the end; read the contact sheet before describing the look.
+set -euo pipefail
+[ $# -eq 2 ] || { echo "uso: analyze.sh <youtube_id> <out_dir>" >&2; exit 2; }
+id="$1"; out="$2"
+[[ "$id" =~ ^[A-Za-z0-9_-]{11}$ ]] || { echo "id inválido: $id" >&2; exit 2; }
+mkdir -p "$out"; cd "$out"
+f="./$id"  # prefixo ./ : ID que começa com hífen não vira opção do ffmpeg
+trap 'rm -f "$f.wav" "$f.mp4"' EXIT
+# Python with librosa + numpy: set VIDEO_STYLES_PY to a venv python, or python3 must have them.
+PY="${VIDEO_STYLES_PY:-python3}"
+"$PY" -c "import librosa, numpy" 2>/dev/null || { echo "precisa de librosa e numpy: pip install librosa numpy (ou VIDEO_STYLES_PY=/caminho/do/python)" >&2; exit 2; }
+for bin in yt-dlp ffmpeg ffprobe; do command -v "$bin" >/dev/null || { echo "falta $bin no PATH" >&2; exit 2; }; done
+url="https://www.youtube.com/shorts/$id"
+fmt="bv*[height<=1280][ext=mp4]+ba[ext=m4a]/b[ext=mp4]/b"
+# When YouTube answers "confirm you're not a bot", the android player client usually still works (no cookies used).
+[ -f "$f.mp4" ] || yt-dlp -q --no-warnings -f "$fmt" --merge-output-format mp4 -o "$f.mp4" "$url" \
+  || yt-dlp -q --no-warnings --extractor-args "youtube:player_client=android" -f "$fmt" --merge-output-format mp4 -o "$f.mp4" "$url"
+meta_fmt="%(view_count)s|%(like_count)s|%(channel)s|%(upload_date)s|%(title)s"
+yt-dlp -q --no-warnings --skip-download --print "$meta_fmt" "$url" > "$f.meta" \
+  || yt-dlp -q --no-warnings --skip-download --extractor-args "youtube:player_client=android" --print "$meta_fmt" "$url" > "$f.meta" \
+  || echo "NA|NA|NA|NA|NA" > "$f.meta"
+ffmpeg -nostats -i "$f.mp4" -vf "select='gt(scene,0.28)',showinfo" -an -f null - 2>&1 | grep -o "pts_time:[0-9.]*" | cut -d: -f2 > "$f.cuts" || true
+ffmpeg -loglevel error -y -i "$f.mp4" -ac 1 -ar 22050 "$f.wav"
+# Video without audio: ebur128 prints no "I:" line; grep would abort under pipefail. Record "NA" instead.
+lufs=$(ffmpeg -i "$f.mp4" -af ebur128 -f null - 2>&1 | grep -E "^\s+I:" | tail -1 | awk '{print $2}' || true)
+[ -n "$lufs" ] || lufs="NA"
+d=$(ffprobe -v error -show_entries format=duration -of csv=p=0 "$f.mp4")
+step=$(python3 -c "print(max(1.0, $d/16))")
+ffmpeg -loglevel error -y -i "$f.mp4" -vf "fps=1/$step,scale=200:-1,tile=8x2" -frames:v 1 "$f-sheet.png"
+"$PY" - "$id" "$lufs" <<'PY'
+import librosa, numpy as np, json, sys
+id, lufs = sys.argv[1], sys.argv[2]
+y, sr = librosa.load(f"{id}.wav", sr=22050); dur = len(y)/sr
+cuts = [float(x) for x in open(f"{id}.cuts").read().split()]
+H, P = librosa.effects.hpss(y, margin=2.0)
+hop = 256
+r = librosa.feature.rms(y=y, hop_length=hop)[0]; rh = librosa.feature.rms(y=H, hop_length=hop)[0]; rp = librosa.feature.rms(y=P, hop_length=hop)[0]
+S = np.abs(librosa.stft(y, hop_length=hop)); f = librosa.fft_frequencies(sr=sr)
+low = S[f < 150].sum(0)/(S.sum(0)+1e-9); cent = librosa.feature.spectral_centroid(S=S, sr=sr)[0]; flat = librosa.feature.spectral_flatness(S=S)[0]
+tfr = librosa.frames_to_time(np.arange(len(r)), sr=sr, hop_length=hop)
+on = librosa.onset.onset_detect(onset_envelope=librosa.onset.onset_strength(y=P, sr=sr, hop_length=hop), sr=sr, hop_length=hop, delta=0.3)
+thr = np.percentile(rp, 80); kinds = {"pop_click":0, "whoosh":0, "impact_low":0}
+events = []
+for o in on:
+    seg = slice(o, min(o+60, len(rp))); e = rp[seg]; pk = e.max()
+    if pk < thr: continue
+    n = (e > pk*0.3).sum()*hop/sr; k = None
+    if low[seg].mean() > 0.30 and n > 0.12: k = "impact_low"
+    elif n < 0.09 and cent[seg].max() > 3000: k = "pop_click"
+    elif n >= 0.15 and flat[seg].mean() > 0.04: k = "whoosh"
+    if k: kinds[k] += 1; events.append([round(float(tfr[o]),2), k])
+on_cut = sum(1 for c in cuts if any(abs(t-c) < 0.15 for t,_ in events))
+shots = np.diff([0]+cuts+[dur]) if dur else np.array([0])
+bed = float(np.percentile(rh,15)/(np.percentile(rh,95)+1e-9))
+meta = open(f"{id}.meta").read().strip().split("|", 4)
+json.dump({"id": id, "url": f"https://www.youtube.com/shorts/{id}", "views": meta[0], "likes": meta[1], "channel": meta[2], "upload_date": meta[3], "title": meta[4],
+  "duration_s": round(dur,1), "cuts": len(cuts), "cuts_per_10s": round(len(cuts)/dur*10,2), "shot_median_s": round(float(np.median(shots)),2), "shot_mean_s": round(float(shots.mean()),2),
+  "music_bed_ratio": round(bed,3), "music_bed_guess": "continua" if bed > 0.13 else ("fraca" if bed > 0.07 else "sem"),
+  "sfx_counts": kinds, "sfx_per_10s": round(sum(kinds.values())/dur*10,2), "sfx_on_cuts": f"{on_cut}/{len(cuts)}", "sfx_events": events[:40],
+  "integrated_lufs": lufs, "note": "sfx/música por heurística (HPSS+onsets); não identifica o NOME do efeito"}, open(f"{id}.json","w"), ensure_ascii=False, indent=1)
+print(f"{id} ok")
+PY

package/src/skills-lib/wizz-video-styles/scripts/build_catalog.py

@@ -0,0 +1,128 @@
+"""Rebuild the catalogue from its data.
+
+data/references.json (every measured short, each with a `style`) +
+data/recipes.json (the qualitative recipe of each style)
+→ data/styles.json (per-style numbers) and references/styles/<style>.md.
+
+Run after adding references (see references/research-method.md):
+    python3 scripts/build_catalog.py
+Stdlib only. A style with a recipe but no measured reference is refused:
+the catalogue never shows a style it cannot back with a measurement.
+"""
+import json
+import os
+import statistics as st
+import sys
+
+ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+DATA = os.path.join(ROOT, "data")
+CARDS = os.path.join(ROOT, "references", "styles")
+
+
+def num(value):
+    try:
+        return float(value)
+    except (TypeError, ValueError):
+        return None
+
+
+def views(ref):
+    return int(ref["views"]) if str(ref.get("views", "")).isdigit() else 0
+
+
+def aggregate(refs):
+    by_style = {}
+    for ref in refs:
+        by_style.setdefault(ref["style"], []).append(ref)
+    out = {}
+    for style, lst in by_style.items():
+        cuts = [c for c in (num(r["ritmo"]["cuts_per_10s"]) for r in lst) if c is not None]
+        shots = [s for s in (num(r["ritmo"]["shot_median_s"]) for r in lst) if s is not None]
+        music = [r["musica"]["guess"] for r in lst]
+        on_cuts = [tuple(map(int, r["sfx"]["on_cuts"].split("/"))) for r in lst if "/" in str(r["sfx"].get("on_cuts", ""))]
+        reach = sorted((views(r) for r in lst), reverse=True)
+        ranked = sorted(lst, key=views, reverse=True)
+        out[style] = {
+            "n": len(lst),
+            "nichos": sorted({r["nicho"] for r in lst}),
+            "langs": sorted({(r.get("idioma") or "")[:2] for r in lst if r.get("idioma")}),
+            "views_max": reach[0],
+            "views_median": int(st.median(reach)),
+            "cuts_per_10s": [round(min(cuts), 2), round(st.median(cuts), 2), round(max(cuts), 2)],
+            "shot_median_s": round(st.median(shots), 2),
+            "music": {k: music.count(k) for k in ("continua", "fraca", "sem")},
+            "sfx_on_cuts": [sum(a for a, _ in on_cuts), sum(b for _, b in on_cuts)],
+            "top": [{"channel": r["channel"], "views": r["views"], "url": r["url"]} for r in ranked[:5]],
+        }
+    return out
+
+
+def card(key, recipe, stats, niches):
+    c = stats["cuts_per_10s"]
+    m = stats["music"]
+    hit, total = stats["sfx_on_cuts"]
+    pct = f"{hit}/{total} ({round(100 * hit / total)}%)" if total else "0/0"
+    where = ", ".join(niches[n] for n in stats["nichos"])
+    lines = [
+        f"# {recipe['title']}", "",
+        f"Measured on {stats['n']} shorts ({', '.join(stats['langs'])}) in: {where}.", "",
+        "## Measured", "",
+        "| Metric | Value |", "|---|---|",
+        f"| Reach (views) | max {stats['views_max']:,} · median {stats['views_median']:,} |",
+        f"| Cuts per 10 s | min {c[0]} · median {c[1]} · max {c[2]} |",
+        f"| Median scene length | {stats['shot_median_s']} s |",
+        f"| Music bed (continuous / faint / none) | {m['continua']} / {m['fraca']} / {m['sem']} |",
+        f"| Cuts with a sound effect on them | {pct} |", "",
+        "Cuts come from scene detection (it undercounts overlays and cards); music and SFX from an audio "
+        "heuristic that classifies type, never names the sound ([../research-method.md](../research-method.md)).", "",
+        "## Material it needs", "", recipe["needs"], "",
+        "## When it fits", "", recipe["fits"], "",
+        "## Recipe", "",
+        "**Look**", "", *[f"- {x}" for x in recipe["look"]], "",
+        f"**Palette:** {recipe['palette']}", "",
+        f"**Typography:** {recipe['type']}", "",
+        "**Motion**", "", *[f"- {x}" for x in recipe["motion"]], "",
+        f"**Captions:** {recipe['captions']}", "",
+        f"**Hook (0 to 3 s):** {recipe['hook']}", "",
+        f"**CTA:** {recipe['cta']}", "",
+        f"**Music:** {recipe['music']}", "",
+        f"**SFX:** {recipe['sfx']}", "",
+        "## Do not", "", *[f"- {x}" for x in recipe["dont"]], "",
+        "## Watch before editing", "",
+        *[f"- [{t['channel']}]({t['url']}) · {views(t):,} views" for t in stats["top"]], "",
+    ]
+    if recipe.get("template"):
+        lines += ["## Template", "",
+                  f"`{recipe['template']}` (Remotion). Copy it into the project's `src/`, set the theme and the scenes; see its README.", ""]
+    return "\n".join(lines)
+
+
+def main():
+    with open(os.path.join(DATA, "references.json"), encoding="utf-8") as fh:
+        refs = json.load(fh)
+    with open(os.path.join(DATA, "recipes.json"), encoding="utf-8") as fh:
+        recipes = json.load(fh)
+    no_views = [r.get("id", "?") for r in refs if not str(r.get("views", "")).isdigit()]
+    if no_views:
+        sys.exit(f"references without numeric views (fix the data before building): {no_views}")
+    bad_niche = sorted({r["nicho"] for r in refs} - set(recipes["niches"]))
+    if bad_niche:
+        sys.exit(f"references with a niche missing from recipes.json niches: {bad_niche}")
+    unknown = sorted({r["style"] for r in refs} - set(recipes["styles"]))
+    if unknown:
+        sys.exit(f"references with a style that has no recipe: {unknown}")
+    stats = aggregate(refs)
+    unbacked = sorted(set(recipes["styles"]) - set(stats))
+    if unbacked:
+        sys.exit(f"recipes with no measured reference: {unbacked}")
+    with open(os.path.join(DATA, "styles.json"), "w", encoding="utf-8") as fh:
+        json.dump(stats, fh, ensure_ascii=False, indent=1)
+    os.makedirs(CARDS, exist_ok=True)
+    for key, recipe in recipes["styles"].items():
+        with open(os.path.join(CARDS, f"{key}.md"), "w", encoding="utf-8") as fh:
+            fh.write(card(key, recipe, stats[key], recipes["niches"]))
+    print(f"{len(refs)} references, {len(stats)} styles, cards in references/styles/")
+
+
+if __name__ == "__main__":
+    main()

package/tools/installer/core/project-instructions.js

@@ -22,8 +22,9 @@ const INSTRUCTION_FILES = ['CLAUDE.md', 'AGENTS.md'];
 // claude-haiku-4-5 continua atual (não há Haiku mais novo em 2026-09-30).
 // `claude-(opus|sonnet)-5` sem sufixo foi trocado por 5-5; o lookahead evita
 // casar o próprio `-5-5`.
+// A família GPT 5.x (Codex) saiu da escada em 2026-10-05: a família GPT-6 assumiu.
 const RETIRED_MODEL_RE =
-  /\bclaude-(?:instant|2|3|3-5|3-7)(?:-[a-z0-9]+)*\b|\bclaude-(?:opus|sonnet)-4(?:-\d+)?\b|\bclaude-(?:opus|sonnet)-5\b(?!-)/gi;
+  /\bclaude-(?:instant|2|3|3-5|3-7)(?:-[a-z0-9]+)*\b|\bclaude-(?:opus|sonnet)-4(?:-\d+)?\b|\bclaude-(?:opus|sonnet)-5\b(?!-)|\bgpt-5(?:\.\d+)?(?:-[a-z]+)*\b/gi;
 
 const LEGACY_RULES = [
   { id: 'gsd', re: /\bGSD\b|get-shit-done|\/gsd[:-]/, why: 'GSD foi substituído pelas fases e checkpoints do Wizz Method' },