@officexapp/vidfarm-devcli 0.21.53 → 0.21.54

package/experimental/sticker-slideshow-tips.md

@@ -0,0 +1,1486 @@
+---
+name: sticker-slideshow-tips
+video_type: Tip slideshow — N still slides on a paper canvas, each one real tip + die-cut cutouts, one slide shilled. Ships as a photo carousel; a 3s-per-slide MP4 is the proof copy (TikTok / Reels / Shorts / Pinterest)
+checks:
+  duration_sec: 15-27       # the VIDEO version. N slides x 3.0s exactly. 5 slides = 15s, 9 = 27s
+  aspect: 9:16
+  first_frame_visual: required
+  first_frame_text: required
+  text_by_sec: 0            # slide 1's headline is up at frame 0. There is no build-on
+  captions: optional        # there is no narration to caption. The slide text IS the content
+  audio: optional           # the carousel is silent; the poster adds sound in-app. A bed on the
+                            # proof MP4 is allowed but never load-bearing — see "Audio DNA"
+  font_regime: required
+  safe_zone: required
+  # CAREFUL: `scenes` counts VISUAL LAYERS, not slides. One slide is 1 background
+  # (photo modes only) + 1-3 cutouts, so a 7-slide deck is 10 layers on paper and 24
+  # with backgrounds. The real slide-count gate is 5-9 SLIDES — under 5 reads as a
+  # single post, over 9 and nobody swipes to the end — and only gate 2 (slides-qa.py,
+  # which counts exported PNGs) can actually enforce it. This range is the loose
+  # upper bound: 9 slides x (1 bg + 3 cutouts).
+  scenes: 5-40
+  max_scene_sec: 3.2        # 3.0 is the number. The 0.2 is render tolerance, not licence
+  # A tip slide carries TWO text runs (headline + subtext) and there are up to 9 slides, so the
+  # whole-composition count is ~18. This is a READ format, not a subtitle format — the
+  # 1-3 ceilings in the sibling harnesses are for videos where a voice is doing the work.
+  max_text_cards: 20
+  max_simultaneous_text: 2  # headline + subtext. A third text run on a slide is clutter
+  max_words_per_cue: 16     # the headline is ONE read at arm's length, 3 lines maximum
+  max_dead_air_sec: 1.0     # there is never a frame without text. A gap means a broken slide
+  max_tail_sec: 0.5
+  forbid_text:
+    - link in bio
+    - sign up
+    - free trial
+    - get started
+    - book a demo
+    - download now
+    - download the app
+    - learn more
+    - swipe up
+    - visit our website
+    # the bait lives in the POST caption, never on a slide:
+    - comment below
+    - save this
+    - follow for more
+---
+
+# Sticker Slideshow — Tips
+
+> A public Vidfarm prompt. Read it end to end before you build.
+
+## The reference deck — watch it first
+
+**Open it before you write a word.**
+
+| | |
+|---|---|
+| **Template** | `template_019fed9bf792762397ca3dae9f49f995` |
+| **Title** | "Cutout Tip Slides" |
+| **Watch it** | <https://vidfarm.cc/discover/templates/feed?template=template_019fed9bf792762397ca3dae9f49f995> |
+| **Fork it** | `vidfarm serve template_019fed9bf792762397ca3dae9f49f995` |
+| **Origin** | tiktok.com |
+| **Shape** | 18.077s · 6 slides · `popularGroup: Slideshows` |
+
+**It runs at exactly 3.0s per slide** — 18.077/6. That is where the 3.0s constant comes from.
+
+## The one test — the three-way lock
+
+Three questions. **All three must be yes** or you have not got a tip slideshow yet.
+
+| # | Question | If no |
+|---|---|---|
+| 1 | **Actionable** — could a stranger do this tonight, without the product, and be better off? | You wrote a fact, or a platitude. Both die |
+| 2 | **The shill is a tip too** — delete the brand name from slide 3. Is what's left still one of the best tips in the deck? | You wrote an ad and put tips around it |
+| 3 | **Still-complete** — freeze any slide as a PNG with no context. Does it stand alone? | You built a video. This format ships as pictures |
+
+Test 3 is the one that gets skipped, and it is structural, not cosmetic: **the slides are the
+master and the video is a rendering of them.** Nothing may carry meaning across a cut, nothing may
+be revealed by motion, no slide may depend on the slide before it. The moment a slide only makes
+sense in sequence, the carousel — which is the real deliverable — is broken, and you will not
+notice, because you were watching the MP4.
+
+---
+
+## Part 0 — who this is for
+
+**A person who already wants the outcome and does not know the mechanics.** Not a buyer. Someone
+saving a list.
+
+Before you write anything, put three lines at the top of your build log:
+
+```
+TOPIC:     Finding the best eats in NYC
+READER:    Someone who lives here, eats out twice a week, and keeps ending up somewhere fine
+OFFER:     dishcover.io — search restaurants by DISH instead of by restaurant
+```
+
+- **The reader is not the buyer, and that is the point.** They are saving a list about eating in
+  NYC. The product reaches them because it is genuinely one of the six answers, not because the
+  deck was aimed at them.
+- **Register: someone who knows the city, telling you what they know.** Flat, certain, no hedging,
+  no "pro tip", no "you might want to consider". A tip with a hedge in it reads as a guess.
+- **They will read at arm's length, at speed, with sound off.** The headline gets one glance. If it
+  needs two, it is a paragraph.
+
+---
+
+## Structural DNA — the deck
+
+**N slides, 3.0 seconds each, no exceptions.** The reference is 6 slides / 18.077s. Five is the
+floor (fewer reads as a single post) and nine is the ceiling (nobody swipes to slide ten).
+
+| Slot | Slide | Job |
+|---|---|---|
+| **1** | **The cover** | Name the list literally + a hero picture. The thumbnail and the promise |
+| **2** | Best tip | Pay the reader immediately. The single strongest thing you know |
+| **3** | Second-best tip | Pay them again, before you ask for anything |
+| **⌈N/2⌉** | **The shill** | The tip that happens to be what the product does. Dead centre |
+| **…** | Tips | Keep the average high. A weak slide here is where people leave |
+| **N** | The tip you'd tell a friend last | The one that earns the save. Not a CTA, not a recap |
+
+**The shill sits in the MIDDLE — slide 4 of 7, slide 3 of 5, slide 5 of 9.** Not a fixed number: a
+fixed number breaks the moment the deck changes length. What is fixed is the *condition* —
+**at least two real tips must be paid before the ad, and at least two after it.**
+
+Why both halves matter. Put the brand earlier and the reader has had one tip, which is not enough to
+have decided the deck is worth their time; they read the ad as the point of the whole thing and
+leave. Put it last and the deck *resolves* into a pitch, which is the same failure the meme harness
+names in Rule 2 — a list that ends on the product retroactively turns every tip before it into
+set-up. The middle is the only slot where the reader is committed and not yet suspicious.
+
+**Counting the cover.** The cover is not a tip, so it does not pay anyone — with a cover, the
+earliest legal shill slot is **slide 4**. A 7-slide deck is cover + 6 tips with the ad at 4: two
+tips in front of it, three behind.
+
+**Exactly one shill in the deck.** Two brand mentions and it stops being a list with a product in
+it and becomes an ad with a list on it. There is no second-mention exception and no "but the last
+slide is subtle".
+
+**Slides do not build.** No "1 of 6" counters, no "and finally", no "but here's the thing" carrying
+into the next slide, no numbered sequence where slide 4 depends on slide 3. Every slide is a
+standalone poster. Shuffle the middle slides at random; if the deck breaks, it was a video.
+
+### Length
+
+`N × 3.0s`, exactly. Set `data-duration` on the root to `N * 3.0` and every slide clip to `3.0`.
+
+**Identical beat lengths are the FORMAT here, not the defect.** The whole-video review pass in
+`references/reviewing-renders.md` flags "N identically-long beats" as a machine-made tell, and it
+is right about every other format in the catalog. It is wrong about this one: a slideshow that
+lingers on slide 3 is a slideshow that is selling. Do not vary the durations to look hand-made.
+Write it in the handoff so the next reviewer does not "fix" it.
+
+---
+
+## The tips — the only thing you actually write
+
+### The grammar
+
+Every headline is **one imperative sentence**, present tense, second person implied.
+
+```
+Never eat on the block with the menus outside.
+Go at 5:30 or go at 9:45.
+Walk two more blocks.
+```
+
+- **Verb first where you can.** "Walk two more blocks" beats "It's worth walking two more blocks".
+- **≤16 words, ≤3 lines.** Break the lines by hand at the sense break, not wherever the box wraps.
+- **No question headlines.** A question defers the payoff to the subtext; the headline is the
+  payoff.
+- **No "tip #4:" prefix, no numbering, no emoji in the headline.**
+- **Two weights in one sentence** carry the emphasis — see *Typography DNA*. That is your only
+  emphasis tool: no colour, no underline, no size change mid-sentence.
+
+**The subtext is the REASON, and it is optional.** One or two lines, smaller, lighter, directly
+under the pictures. It answers "why" and never repeats the headline in other words.
+
+**The shill slide, and only the shill slide, may run to three lines.** It has one job the others do
+not — naming the offer — and squeezing that into two lines is what tempts you into cutting the
+reason down to a slogan. Two lines of tip, one line naming the tool. Every other slide stays at two:
+a three-line subtext elsewhere means the tip needed a paragraph, which means it was not a tip.
+
+```
+headline:  Trust a 30-item menu over a 200-item one.
+subtext:   A short menu means a kitchen that repeats. A long one means a freezer.
+```
+
+A subtext that restates the headline is worse than no subtext — it tells the reader the deck has
+nothing more to say, and they stop reading the subtexts for the rest of the carousel.
+
+### The specificity ladder — the single biggest lever
+
+Same tip, four rungs. Rung 1 gets scrolled. Rung 4 gets saved.
+
+| Rung | Line | Why |
+|---|---|---|
+| 1 | Do your research before you go out | True of everything. Says nothing |
+| 2 | Avoid tourist-trap restaurants | Names a category the reader already avoids |
+| 3 | Walk two blocks further from the train | A real instruction, but no number to hold |
+| 4 | **Every train exit has a trap in the first 200 feet. Walk past it.** | A mechanism + a number. This is the rung |
+
+**Rung 4 is a claim you can be wrong about.** That is what makes it worth reading — and it is why
+Rule 6 exists: a rung-4 number has to be a number you'd defend, not one you rounded into place.
+
+### Banned tips
+
+Each of these fails in a way that is invisible while you're writing it:
+
+- **The tip that is really the product.** "Use an app that lets you search by dish" on slide 5 is a
+  second shill wearing a hat. One shill, and it is slide 3.
+- **The tip nobody disputes.** "Read reviews before you go", "book ahead on weekends". Being true
+  is not the bar; being *told* is.
+- **The tip that needs the reader to already be an expert.** If the instruction assumes they can
+  read a menu in Cantonese, they were never the reader.
+- **The tip that is a preference.** "Sit at the bar" is a preference. "Sit at the bar — the bar
+  gets the pass first and the food is hotter" is a mechanism.
+- **The tip that costs money to follow.** A list of things a reader can do for free travels; a list
+  of things they can do for $80 is an ad for restaurants.
+- **The hedged tip.** "You might want to try…", "it's often worth…", "consider…". Cut the hedge and
+  the tip either stands up or was never a tip.
+
+### The cold-viewer test
+
+Cover the brand name on slide 3. Read all N slides. **Does the deck still read as a good list?**
+If slide 3 is now visibly the weakest one, you did not write a tip — you wrote a placement and
+dressed it. Rewrite slide 3, not the deck.
+
+---
+
+## The shill slide
+
+The one that pays for the work. It obeys every rule above **plus** five of its own.
+
+```
+headline:  Search the dish, not the restaurant.
+subtext:   "Best birria in Queens" finds you dinner. "Best Mexican near me" finds you a parking lot.
+           dishcover.io is built to search that way.
+cutouts:   a bowl of birria (photographic), a hand-drawn arrow pointing at it
+```
+
+1. **It is a mechanism, never an outcome.** "Search by dish" is a mechanism. "Find better food" is
+   an outcome, and an outcome is a promise the deck cannot keep.
+2. **It looks identical to every other slide.** Same canvas, same type, same weights, same layout
+   grid, same number of cutouts. **No accent colour, no logo, no screenshot, no border.** The
+   moment slide 3 is visibly styled differently, the reader's eye finds the ad before they read it.
+3. **The brand is named once, lowercase, in the subtext.** Never in the headline — a headline with
+   a brand in it is a banner. `dishcover.io`, not `Dishcover™`, not `Dishcover.io (free!)`.
+4. **The tip works without the product.** A reader who never taps anything can still go search
+   `"best birria in queens"` in a normal search box and be better off. That is what makes the
+   slide honest, and it is exactly why it converts: you demonstrated the mechanism instead of
+   describing a feature.
+5. **No competitor is named doing something badly.** "Best Mexican near me" is a *query shape*, not
+   a product. Naming a real competitor in a deck like this is a legal problem and a credibility
+   problem in the same sentence.
+
+**Where the shill goes when the product is not a search tool.** The rule generalises: find the one
+slide where the honest, product-free version of the tip is *the thing the product automates*, and
+put the product name in that slide's subtext as the tool that does it. If no such tip exists, the
+product does not fit this format — use `meme-recaption` or `ugc-reaction-greenscreen`.
+
+---
+
+## Casting the pictures
+
+### The art class — pick one, hold it across the whole deck
+
+**The loudest possible failure in this format is mixed art classes.** A photographic cutout on
+slide 1, a flat vector icon on slide 2 and a 3D render on slide 4 reads as three different people
+built the deck. Pick one class in the build log before you source anything:
+
+| Class | What it is | Best for |
+|---|---|---|
+| **Photographic cutout** *(the reference, and the default)* | A real photo, die-cut, no background, soft contact shadow | Anything you can eat, wear, hold or stand in front of |
+| **Flat sticker** | IconScout `--style sticker` — thick-outline, limited palette | Abstract tips with no physical noun (time, money, queues) |
+| **Hand-drawn mark** | Felt-tip arrow, circle, cross, underline | The *annotation* layer only — see below |
+
+**Hand-drawn marks are not an art class, they are punctuation.** An arrow between two cutouts, a
+cross over the wrong option, a circle around the right one. They may appear alongside either of the
+first two classes, and they should — the reference's whole slide is `[crushed bottle] → [nice
+bottle]`, and the arrow is what makes it a sentence instead of two objects.
+
+### The picture must be the tip's noun
+
+The cutout is not decoration and it is not a mood. **It is the thing the sentence is about.** If
+the tip says "a 30-item menu beats a 200-item one", the cutouts are two menus of visibly different
+length. If the sentence has no physical noun, either find the noun ("go at 5:30" → a clock face) or
+**give the slide no picture at all** — an all-type slide inside a picture deck is a legitimate beat
+and infinitely better than a stock photo of a smiling person that means nothing.
+
+Test it the way the meme harness tests enactment: **cover the text. Can a stranger guess the
+tip within about two seconds?** Not exactly — approximately. If the pictures could belong to any of
+your six slides, they belong to none of them.
+
+### Sourcing, cheapest first
+
+The ladder, and stop at the first rung that answers the slide:
+
+```bash
+# 1. FREE designer stickers/illustrations — no key, no account, search is free
+vidfarm iconscout "bowl of birria"  --style sticker --free
+vidfarm iconscout "alarm clock"     --style sticker --free
+vidfarm iconscout get <uuid> --format png --size 1024 --out media/cutouts/clock.png
+
+# 2. FREE stock photo -> die-cut it locally for $0 (ONNX matting, no plate needed)
+vidfarm media search "birria tacos overhead" --type image
+vidfarm remove-background ./birria.jpg --out media/cutouts/birria.png
+
+# 3. Art the director already owns — lift an element out of their own site/menu/screenshot
+vidfarm mask ./their-site.png --crop 120,340,600,600 --out media/cutouts/dish.png
+
+# 4. ONE generated sheet for the WHOLE deck — one billed image job, one art style
+vidfarm sticker-pack --generate "die-cut photographic food cutouts on a flat #FF00FF plate" \
+  --items "birria bowl,laminated menu,alarm clock,subway sign" --out-dir media/cutouts
+```
+
+**Rung 4 is the one to reach for in `hybrid`, and reach for it ONCE.** One sheet holding every
+cutout in the deck costs one image job and — far more important — guarantees one art class. N
+separate cutout jobs cost N times as much *and* produce N slightly different styles, which is the
+failure this section opens with.
+
+**In `minimize` + `interactive`, rung 4 is free.** Hand the user one prompt, get one sheet back,
+split it locally:
+
+```bash
+vidfarm handoff image --theme "die-cut photographic food cutouts, NYC eating deck" \
+  --items "birria bowl,laminated menu board,alarm clock,subway exit sign,short menu,long menu"
+# they paste it into meta.ai / free ChatGPT / a HuggingFace Space, hand back sheet.png
+vidfarm sticker-pack ./sheet.png --items "birria bowl,laminated menu board,alarm clock,…" \
+  --out-dir media/cutouts
+```
+
+### ⚠️ Attribution — the one that bites this format
+
+**IconScout free assets and most free stock require a credit line, and a credit line on the frame
+is brand chrome, which is banned.** This format has an out the sibling harnesses do not: a
+carousel always ships with a **post caption**. Put every credit there, in the caption, one line at
+the bottom. Never on a slide, never on a "credits" slide N+1, never in 12px grey in a corner.
+
+If a licence genuinely requires the credit to be *on the image*, do not use that asset — go back to
+rung 2 (matted stock, no attribution) or rung 4 (generated).
+
+### The alpha has to actually be clean
+
+A cutout on a near-white canvas shows every defect a cutout on footage would hide. **A grey halo,
+a soft 40%-alpha edge, or a leftover rectangle of background is instantly visible on white and
+invisible in a thumbnail.** Check it, don't trust it:
+
+```python
+# any cutout: how much of it is actually opaque?
+a = alpha_channel(png)
+print((a == 0).mean(), ((a > 0) & (a < 255)).mean(), (a == 255).mean())
+# healthy: mostly 0 or 255, a percent or two in between.
+# tens of percent "partial" = a bad key, and on white it WILL show as a grey box.
+```
+
+Match the key mode to the shape — this is the same trap the greenscreen harness documents, and it
+costs a wasted job every time it is rediscovered:
+
+| Art | Key mode | Why |
+|---|---|---|
+| **Hollow** — an arrow outline, a ring, a frame | `--key-mode flat` | the interior never touches the frame edge, so a connectivity keyer *keeps* it and you get a filled blob |
+| **Solid** — a photographic dish, a bottle, a person | `--key-mode smart` *(default)* | flat mode's soft alpha ramp eats a photograph — measured at 65% partial alpha on a first attempt vs 0.6% in smart mode |
+
+And pick the plate colour **against the subject**: a green plate behind green food keys holes
+through the food. Magenta `#FF00FF` is the safe default for anything edible.
+
+---
+
+## Visual DNA
+
+### The canvas
+
+**Near-white paper, and it never changes.** Sampled off the reference: `#FDFDFD`. Use
+`#FDFCFA` (a hair warm, kinder on OLED) or plain `#FFFFFF`. Text is near-black `#111111`, not pure
+`#000000`.
+
+- **One canvas colour for the whole deck.** Not one per slide, not a slow tint drift, not a
+  different background for the shill.
+- **No footage, ever.** No photo backgrounds, no blurred plate, no gradient, no texture, no
+  drop-shadowed cards. The page is a page.
+- **No border, no frame, no margin rule, no page number.**
+- **This is deliberately not the sibling harnesses' look.** Those put text on footage and need a
+  stroke to survive it. Here the contrast is already maximal, and a stroke or a text shadow on a
+  white page is exactly the thing that makes a slide read as machine-made. **No stroke, no text
+  shadow, no plate behind the type.** The only shadow anywhere in the deck is a soft contact
+  shadow under a photographic cutout, and even that is optional.
+
+### Canvas DNA — four background modes
+
+The reference is a white page, and that is still the default. But a photograph behind the type is
+the other half of this format in the wild, and it changes one thing structurally: **on a page the
+type needs nothing, on a photo the type needs help.** Pick a mode per deck — never per slide — and
+the whole legibility contract comes with it.
+
+| Mode | The picture | Ink | Reads as |
+|---|---|---|---|
+| **`paper`** *(default)* | none — flat `#FDFCFA` | `#111` | Editorial, clean, the reference. The cutouts are the only colour |
+| **`photo-darken`** | blurred, darkened | **white** | Moody, premium, night-out. **The most legible of the three** |
+| **`photo-lighten`** | blurred, brightened, desaturated | `#111` | Airy, daylight, magazine. Closest to `paper` with context |
+| **`photo-gray`** | blurred, greyscale, neutral wash | **white** | Editorial monochrome. **The tightest on contrast — see below** |
+
+**A photo mode is not decoration, it is context.** Each slide's background is the *place* the tip is
+about — the storefront, the dining room at service, the subway platform. It is blurred hard enough
+that it is atmosphere, never a picture competing with the cutout in front of it. If you cannot name
+what a background is doing for its slide, use `paper`.
+
+**Backgrounds are free.** `vidfarm media search "<place>" --type image` (pixabay/openverse) covers
+this entirely at $0 — no generation, in any cost mode.
+
+#### The legibility contract
+
+Each mode is a *set* of values tuned together — photo filter, wash, ink colour, text shadow, cutout
+shadow. **Changing one alone is how a deck ends up with a 3:1 headline that looked fine on the one
+slide you opened.** They live in `MODES` in Appendix A.
+
+- **The blur and the dim are DIFFERENT ELEMENTS.** A single backdrop that is both blurred and dimmed
+  is the website-modal staging pattern, and `vidfarm qa` is right to call it (`modal-scrim`). Blur
+  the photograph on the `<img>`; carry the dim on a separate flat colour layer above it. Identical
+  result, and it is not modal staging.
+- **The wash is a BANDED GRADIENT, not a flat tint.** A flat wash has to be deep enough for the
+  worst band, which flattens the whole picture into mud. A gradient that is strong across the
+  headline band (0–32%) and the subtext band (72–100%) and weak through the cutout band (44–64%)
+  buys the contrast exactly where the text is and lets the photograph breathe where it doesn't.
+- **The photo is scaled to 110% and offset −5%.** A blurred image at exactly 100% shows a pale
+  vignette on all four edges — the blur sampling past the frame — and it reads as a rendering bug.
+- **Rule 5 inverts on a photo.** On `paper`, a text shadow is decoration and banned. On a photo it
+  is legibility, so `photo-*` modes carry a soft diffuse shadow (no stroke, still no plate, still no
+  card). The cutouts get a real contact shadow for the same reason: a die-cut object with no shadow
+  floats unconvincingly over a photograph.
+- **Hand-drawn MARKS follow the ink; photographs never do.** A `#111` arrow on `photo-darken` is
+  invisible — and invisible in precisely the spot the eye is being pointed. Write `{ink}` into the
+  mark's filename (`arrow-{ink}.svg`) and the builder resolves `arrow-dark.svg` / `arrow-light.svg`
+  per mode. It is opt-in on purpose: recolouring a photographic cutout would ruin it.
+
+#### Prove the text is readable — measure it, don't look at it
+
+**This is the check that makes photo modes safe to ship**, and it cannot be done by eye: a headline
+that reads perfectly on slide 2's dark doorway can dissolve on slide 5's bright tiled wall, and you
+only find out on the slide you did not open.
+
+`slides-qa.py --ink light|dark` (Appendix C) recovers each text band's background by **blurring the
+band hard enough to erase the glyphs**, takes the worst-case patch of what is left, and computes the
+WCAG contrast ratio against the mode's known ink colour.
+
+| Ratio | Verdict |
+|---|---|
+| **< 4.5:1** | **Fail.** Deepen the wash, or switch polarity (`darken` ↔ `lighten`) |
+| 4.5–7:1 | Passes AA. Fine for 70px display type; tighten if you can |
+| **≥ 7:1** | Ship it |
+
+Measured on the reference build: `photo-darken` and `photo-lighten` clear 7:1 on every band of every
+slide. **`photo-gray` lands at 6.1–6.9:1 and is the mode with the least headroom** — greyscale
+throws away the colour separation the other two lean on, so it is squeezed between "picture still
+visible" and "white text still readable". It is legal and it looks good; it is just the one mode
+where a bright new background can push a slide under, so re-measure whenever you swap its photos.
+
+**Three checks in `slides-qa.py` are `paper`-only** and are skipped automatically once `--ink` is
+set: canvas-drift (a photo deck is *supposed* to change picture every slide), headline-band
+levelling (the detector locks onto photo detail instead of glyphs), and reserved-zone ink (every
+full-bleed photograph has ink everywhere). On a photo deck the contrast gate replaces all three.
+
+### The layout grid
+
+Three horizontal bands, held identically on every slide. Read off the reference and expressed as
+percentages of the 1080×1920 canvas:
+
+```
+  0% ─────────────────────────────────────
+        (air — nothing lives here)
+ 14% ┌───────────────────────────────────┐
+     │  HEADLINE   centred, 1-3 lines    │
+ 30% └───────────────────────────────────┘
+        (air)
+ 34% ┌───────────────────────────────────┐
+     │                                   │
+     │  CUTOUTS   1-3 objects + marks    │
+     │                                   │
+ 66% └───────────────────────────────────┘
+        (air)
+ 73% ┌───────────────────────────────────┐
+     │  SUBTEXT   centred, 1-2 lines     │
+ 80% └───────────────────────────────────┘
+        (air — nothing lives here)
+100% ─────────────────────────────────────
+```
+
+- **The bands do not move between slides.** A headline that is 2 lines on slide 1 and 3 lines on
+  slide 2 still starts at 14%. Vertically centring each slide's content individually is the single
+  most common way a deck acquires a "wobble" nobody can name — and it is invisible slide by slide.
+- **The bottom 20% and the top 14% stay empty.** That is where the platform puts the caption, the
+  username, the buttons, and the swipe affordance.
+- **Everything is centred.** The headline block centres, the cutout group centres, the subtext
+  centres. The reference does exactly this and it is why the page reads as a page.
+
+### Sticker economy — generous, but composed
+
+The user asked for generous sticker use, and this format supports it in a way the video formats do
+not: **a still frame can hold three objects, because the reader controls how long they look.**
+
+- **1–3 cutouts per slide.** One hero object; two for a comparison; three only when they are a
+  *set* (three menu lengths, three times of day).
+- **Plus one hand-drawn mark** — an arrow, a cross, a circle. Marks are not counted in the 1–3.
+- **Slides with zero cutouts are legal and should exist.** One or two all-type slides in a six-slide
+  deck give the eye somewhere to rest, and they are the right answer for a tip with no noun.
+- **Rotate them slightly.** ±3–8° on each object, different per object. Perfectly axis-aligned
+  cutouts on a white page read as a PowerPoint; a slight tilt reads as things laid on a table.
+- **They may overlap the band edges but never each other's silhouettes.** Two cutouts touching read
+  as one confusing object.
+- **Consistent scale relationship.** If a bottle and a person are both on a slide, the person is
+  bigger. Sounds obvious; gets broken every time the two cutouts came from different sources.
+- **One light direction across the whole deck.** If the birria is lit from the upper left, so is
+  the menu board. Contact shadows all fall the same way or you drop them entirely.
+
+### The comparison layout — the format's strongest slide
+
+The reference is one: `[crushed plastic bottle] → [nice steel bottle]`. Two cutouts, an arrow
+between them, and the tip is legible before a word is read.
+
+Use it for any tip of the shape *stop doing X, do Y*. Left object is the wrong one (and it may be
+visibly worse — crushed, greasy, cluttered); right object is the right one; the arrow is
+hand-drawn and points right. **Do not put a cross over the left object AND an arrow** — one mark
+per slide.
+
+---
+
+## Typography DNA
+
+### The regime
+
+**Two weights of ONE family, self-hosted, no fallback chain.** A fallback chain means a missing
+font silently renders as a different font and nothing tells you — the deck just quietly stops
+looking like itself.
+
+| Role | Face | Weight | Size | Line height |
+|---|---|---|---|---|
+| Headline — emphasis | TikTok Sans *or* Montserrat | **800** | 76px | 1.22 |
+| Headline — body | same family | **500** | 76px | 1.22 |
+| Subtext | same family | **500** | 40px | 1.34 |
+
+`vidfarm qa`'s `font-regime` rule errors on Inter / Roboto / Arial / system-ui and warns on any
+family outside the imported regime (Montserrat, TikTok Sans, Abel, Source Code Pro, Yesteryear).
+**Montserrat is the closest match to the reference** and is in the regime.
+
+> **Repo note:** `work/*/media/fonts/` currently carries `TikTokSans-700.woff2` and
+> `TikTokSans-900.woff2` only. This format needs a **mid weight (400–500)** for the non-emphasised
+> half of the headline. Fetch it once and copy it beside the others before you build — do not
+> approximate a 500 by setting `opacity` on a 700, which produces grey text, not light text.
+
+### Mixed weight inside one sentence — the whole voice
+
+This is the reference's signature and it is not decoration. The **800** carries the words that make
+the tip, the **500** carries the connective tissue:
+
+```html
+<b>Stop</b> buying <b>plastic bottle</b><br>
+every day <b>bring your own</b><br>
+at school/work
+```
+
+- **Emphasise the verb and the noun.** Not the adjectives, not "every day", not "the".
+- **2–4 emphasised runs per headline.** One is a typo; five is a highlighter.
+- **The emphasis is the same colour as the body.** Weight only. A coloured word on a white page is
+  a link, and nothing in a video is clickable.
+- **Never emphasise the brand name** on slide 3. It rides in the 500 with everything else — that
+  is the difference between a tip that mentions a tool and an ad.
+
+### Spacing
+
+Getting this wrong is what makes a slide look "off" in a way nobody can point at:
+
+- **Line height 1.22 on the headline.** Tighter reads as a logo; looser reads as a document.
+- **Hand-break the lines at the sense break.** `every day / bring your own` — never let the box
+  decide. Aim for near-equal line lengths so the centred block has a shallow ragged edge.
+- **Letter-spacing `-0.01em` on the headline**, `0` on the subtext. Large type needs the pull-in;
+  small type does not.
+- **Max headline width 84% of canvas.** Full-bleed type is the `caption-oversize` failure and it
+  forces wraps you did not choose.
+- **The gap between the headline block and the cutouts is fixed**, not proportional to how many
+  lines the headline took. This is the band grid again, and it is the same failure.
+
+---
+
+## Motion DNA
+
+**The default is no motion at all.** The reference has none. A deck of still slides, hard cut every
+3.0s, is the correct build and it is what the carousel exports cleanly from.
+
+What is allowed, and only in the MP4:
+
+- **A cutout entrance**, if and only if it **completes within 0.5s** of the slide's start. The
+  still export is taken at slide start **+1.5s** for exactly this reason: the settled state is the
+  slide, and the entrance is a courtesy to the video viewer.
+- **Nothing on the headline or subtext.** Text is up at the slide's frame 0, full opacity, final
+  position. A headline that types on, fades up, or pops per word cannot be exported as a PNG and
+  cannot be read in 3 seconds.
+
+What is banned:
+
+- **Ken Burns.** It is the right tool for a slideshow of *photographs*; it is wrong here, because
+  the canvas is a page and a drifting page is a rendering artifact.
+- **Any transition between slides.** No crossfade, no whip, no slide-in. **Hard butt cuts only** —
+  a carousel swipes, it does not dissolve, and the MP4 should feel like a swipe.
+- **An entrance transition on slide 1.** `vidfarm qa` errors on this (`thumbnail-fade-in`) and it
+  is right: frame 0 is the thumbnail and the first slide of the carousel.
+- **Anything that reveals.** If motion tells the reader something the still does not, the carousel
+  is broken. See test 3.
+
+---
+
+## Audio DNA
+
+**The carousel is silent.** There is nothing to author.
+
+**The MP4 may carry a music bed and nothing else.** No voiceover — a voice reading the slides makes
+the video the master and the slides the leftovers, which inverts the format. If you add a bed:
+
+- Free and keyless: `vidfarm media search "<mood>" --type bgm` (Openverse, CC/CC0, no key).
+- It is **never load-bearing.** No beat-matched cuts, no drop on slide 3. The cuts land on 3.0s
+  because that is the format, not because the music said so.
+- Ship the MP4 **silent as well** if the proof will be judged on the slides. `export-versions.py`
+  in the greenscreen harness does this shape already.
+
+The poster adds a trending sound in the platform's own editor. Say so in the handoff.
+
+---
+
+## The two deliverables
+
+**One composition. Two exports. The slides are the master.**
+
+| Deliverable | What it is | How it is made | Where it goes |
+|---|---|---|---|
+| 📸 **The slideshow** *(the real one)* | N PNGs, 1080×1920, in order | `vidfarm stills` at each slide's start+1.5s | Durable My Files → N ordered URLs → posted as a photo carousel |
+| 🎬 **The video** *(the proof)* | One MP4, N×3.0s, hard cuts | `vidfarm render` on the same composition | The marketplace proof; also postable to a video surface |
+
+### ⚠️ Name the file BOTH ways before you export anything
+
+The two halves of the toolchain disagree about the filename, and each one fails with an error that
+sounds like a missing project rather than a missing alias:
+
+| Tool | Wants | Error when it's absent |
+|---|---|---|
+| `vidfarm stills` / `vidfarm qa` / `vidfarm render` | `composition.html` | *"No composition.html inside … Run `vidfarm pull` first"* |
+| `vidfarm hf …` (the hyperframes passthrough) | `index.html` | *"No composition found … Run npx hyperframes init"* |
+
+Write `composition.html` and symlink `index.html` beside it. One line, and it removes an entire
+class of dead end:
+
+```bash
+ln -sf composition.html ./work/index.html
+```
+
+### Exporting the slides
+
+```bash
+# 6 slides at 3.0s -> settled-state stills at 1.5, 4.5, 7.5, 10.5, 13.5, 16.5
+vidfarm stills ./work --at 1.5,4.5,7.5,10.5,13.5,16.5 --out ./slides --sheet
+
+# If `stills` dies with "Cannot find package '@hyperframes/producer'", the devcli's
+# bundled engine is incomplete on this machine. Do NOT go hunting — the passthrough
+# is the same renderer and it works:
+vidfarm hf snapshot ./work --at 1.5,4.5,7.5,10.5,13.5,16.5 --no-end \
+  -o ./slides --describe false
+# and for the MP4, likewise: vidfarm hf render ./work --output ./work/final.mp4
+```
+
+`--sheet` is not optional here. The contact sheet **is** the carousel, laid out flat, and it is the
+only view in which the sequence-level defects this format is prone to are visible: a wobbling
+headline band, three type sizes, one cutout at the wrong scale, an art class that drifted by slide
+5. Read it as one image before you upload anything. Appendix B automates the whole export.
+
+### Uploading and ordering
+
+The slides must live in **durable** My Files, not the 30-day temp store — a carousel that 404s in a
+month is worse than one that never shipped.
+
+```bash
+for i in 1 2 3 4 5 6; do
+  vidfarm put-file ./slides/slide-0$i.png --folder decks/nyc-eats \
+    --notes "slide $i of 6 — NYC eats tip deck"
+done
+vidfarm files --search "NYC eats tip deck"     # confirm all six, in order
+```
+
+Write the ordered URLs to `slides.json` beside the composition. **Order is data, not a filename
+convention** — a client posting the carousel will paste them in the order you give, and slide 3
+being the shill is load-bearing.
+
+### Submitting to the marketplace
+
+A gig proof has to be **playable**, which is why the MP4 exists. The slides travel beside it:
+
+```bash
+# 1. render the video version
+vidfarm render ./work --wait                       # -> final.mp4
+
+# 2. drop BOTH into the client's upload link, in your own subfolder
+vidfarm shared put <upload-link> ./final.mp4      --subfolder <your-handle>/nyc-eats
+vidfarm shared put <upload-link> ./slides.json    --subfolder <your-handle>/nyc-eats
+for f in ./slides/slide-*.png; do
+  vidfarm shared put <upload-link> "$f" --subfolder <your-handle>/nyc-eats/slides
+done
+
+# 3. submit the MP4 as the proof
+vidfarm gigs submit <gig-id> --task <task-id> --proof https://…/final.mp4
+```
+
+**Say in the submission that the deliverable is the carousel and the MP4 is the preview.** A client
+who receives an MP4 with no note assumes the MP4 is the product, posts it as a video, and gets a
+fraction of the reach the format is for.
+
+---
+
+## The rules
+
+Every rule has its reason attached. A rule without its reason gets argued away by the next agent
+that reads this file.
+
+### Rule 1 — one shill, on slide 3, written as a tip
+
+Any other placement or count changes what the deck *is*. Slide 1 makes it an ad, slide N makes it a
+pitch, two mentions make it a brochure. The shill is in the middle because that is where the reader
+is already paid and not yet suspicious. **Enforced:** the cold-viewer test, and `slides-qa.py`
+checks that exactly one slide contains the brand token.
+
+### Rule 2 — every slide is a complete still
+
+Nothing crosses a cut. No motion carries meaning. The PNG export at start+1.5s **is** the slide, so
+if the PNG is missing something, the slide is missing it. This is the rule the whole format hangs
+on, because the carousel is the deliverable and the video is not.
+
+### Rule 3 — the layout bands never move
+
+Headline at 14%, cutouts 34–66%, subtext at 73%, on every slide, whatever the content length. Per
+slide it looks fine either way; across the deck, individually-centred content produces a vertical
+wobble that reads as sloppy and that nobody can name while looking at one slide.
+
+### Rule 4 — one art class, one light direction, one canvas colour
+
+The fastest way for a deck to look like it was assembled by a machine from four sources is for it
+to have been assembled by a machine from four sources. Source the whole deck in one pass, ideally
+from one sheet.
+
+### Rule 5 — no stroke, no text shadow, no plate, no card
+
+On footage those are legibility. On a white page they are decoration, and `vidfarm qa` will call
+the plate a `card-panel`. Black on white is already the highest contrast available.
+
+### Rule 6 — every number on a slide is a number you would defend
+
+"200 feet", "30-item menu", "5:30". Rung-4 specificity is what makes these tips worth saving, and
+it only works if the numbers are real. A number invented to sound specific is the one thing in the
+deck a reader can catch you on, and a reader who catches one stops believing the other five.
+
+### Rule 7 — no competitor is named doing something badly
+
+Name query shapes, habits, and categories. Never a real product. It is a legal exposure and a
+credibility loss in the same line.
+
+### Rule 8 — credits go in the post caption, never on a slide
+
+An attribution line on the frame is brand chrome, and this format bans brand chrome for the same
+reason the others do. The carousel always has a caption; use it. If a licence forces on-image
+credit, change the asset.
+
+### Rule 9 — the bait is in the post caption
+
+The deck ends on a tip. The comment prompt, the save prompt and the follow prompt all live in the
+caption where the platform expects them. A slide that says "save this" spends the slot that could
+have held the tip that made them save it.
+
+### Rule 10 — production floor
+
+1080×1920, 30fps, `N × 3.0s` exactly, hard cuts, one canvas, no letterboxing, text at frame 0 of
+every slide, `data-duration` on the root equal to the sum.
+
+---
+
+## Cost-saving mode — the whole deck at $0
+
+This format is the cheapest in the catalog because it has no footage. In `minimize` the only thing
+that can cost money is a cutout, and every cutout has a free rung.
+
+| Ingredient | `minimize` ($0) | `hybrid` (~$0.05) |
+|---|---|---|
+| Canvas + type | Hand-authored HTML. Free, forever | same |
+| Cutouts | `iconscout --free` → `media search --type image` + `remove-background` → `mask` | one `sticker-pack --generate` sheet for the whole deck |
+| Fonts | Self-hosted woff2, already on disk | same |
+| Slides export | `vidfarm stills` — local, in-process | same |
+| Video render | `vidfarm render` local | same |
+| Music bed | `media search --type bgm` (Openverse, keyless) | same |
+| **Total** | **$0** | **~$0.05, one image job** |
+
+**In `minimize` + `interactive`, one round trip buys the whole deck's art.** `vidfarm handoff image`
+→ the user runs it in a free web generator → `vidfarm sticker-pack` splits it. Free frontier image
+models are typically better per image than an API budget buys, so this is not the compromise rung —
+it is frequently the best-looking one.
+
+**Claude Code has no image model.** `vidfarm agent-image` is free custom art on Antigravity /
+Gemini CLI / Codex, and it is genuinely $0 in `minimize` there. In a Claude session it is not
+available; drop to the handoff rung and say so rather than pretending.
+
+---
+
+## Bulk generation — this format is built for it
+
+A deck is a manifest plus a template. Once one is built, the tenth costs minutes.
+
+- **Vary the TOPIC, hold the structure.** "Best eats in NYC" → "…in Austin" → "…on a $15 budget" →
+  "…that are open past midnight". Same six slots, same layout, same art class, new tips.
+- **Never reuse a tip across decks in the same account.** A reader who follows you sees both.
+- **The shill slide is the one thing that may repeat verbatim** across a campaign — it is the
+  product's mechanism, and it does not get better by being reworded. Hold it constant and vary
+  everything else; that also makes it the clean variable when you want to test it.
+- **`vidfarm dedupe` before a re-post**, not before a first post. The MP4 is what a platform
+  fingerprints; the slides are not.
+- **Batch the QA.** `vidfarm qa ./decks/* --json` grades N compositions in one pass; Appendix E's
+  `slides-qa.py` runs on the exported PNGs, which is where this format's real defects live.
+
+Sizing a round, capacity, and the ledger belong to `vidfarm experiment`, not to this file.
+
+---
+
+## Quality gates — what is ENFORCED, and by what
+
+| Gate | Tool | What it can actually see |
+|---|---|---|
+| **1. Composition** | `vidfarm qa ./work --harness ./sticker-slideshow-tips.md` | The DOM: font regime, slop, safe zone, frame 0, the `checks:` block above |
+| **2. The stills** | `slides-qa.py ./slides [--ink light\|dark]` *(Appendix C)* | The exported PNGs: blank slides, canvas drift, band alignment, cutout alpha, the one-shill count |
+| **3. The sequence** | **You**, on `slides/contact-sheet.png` | Whether it reads as one deck. Nothing else can see this |
+| **4. The tips** | **You**, cold | Whether any of it is worth saving. No tool has an opinion here |
+
+Gate 1 is free and instant and does not look at pixels. Gate 3 is the one that catches what
+actually ships broken, and it cannot be automated — every first-pass deck in a batch has a
+sequence-level defect its own author called "verified, looks good."
+
+### Two findings that are EXPECTED on this format — do not "fix" them
+
+Both are correct rules firing on a format they were not written for. Record your answer once and
+move on; chasing either one damages the deck.
+
+- **`wall-of-text` on the subtexts** (built-in, gate 1). It warns at 14+ words in one static text
+  layer, on the reasoning that a scrolling viewer reads none of it. **That reasoning does not hold
+  here** — a carousel reader sets their own dwell time, and the reason line is the thing they came
+  for. The fix it suggests (page it into 3–5-word kinetic cues) would destroy the format. Expect it
+  on any subtext over ~13 words and leave them alone.
+- **`partial alpha` on a thin-structured cutout** (gate 2). The 8% threshold is calibrated for solid
+  blobs. An object made of thin members — a sign on a stand, a railing, cutlery, a chair — has far
+  more edge per unit area and lands at 8–12% while being visually perfect. **Judge it on the
+  pixels, not the number:** crop the cutout out of the rendered slide and look for a grey rim. A
+  real halo is obvious at a glance; antialiasing on thin metal is not.
+
+### The failure this format actually ships — the invisible layout collision
+
+Worth stating because it survived a clean render, a clean export and a human glance at one slide:
+`.clip { inset: 0 }` must be reset per layer, that reset lives in the **inline** style attribute,
+and an inline `inset:auto` **beats a `top:` in your stylesheet**. Put the band geometry in a CSS
+class and every text layer silently resolves to `top:auto` — headline and subtext stack on top of
+each other at the head of the frame. It renders, it exports, it passes gate 1, and it is unusable.
+**Geometry inline, typography in the class.** Gate 2's *"no headline found in the band"* is the
+check that catches it.
+
+---
+
+## Pre-flight checklist
+
+Run this before you build, and answer it honestly after. An unchecked box is a rewrite, not a fix
+in the edit.
+
+**The deck**
+
+- [ ] Slide 1 is a COVER and says "Tips for…" (or equivalent) literally — no clever hook.
+- [ ] The cover carries a hero picture and a NUMBER, and there is only one cover.
+- [ ] The topic is ALSO in the post caption, which is where the bait and any credits live.
+- [ ] N is between 5 and 9 (cover included), and every slide is exactly 3.0s.
+- [ ] Slide 2 is the strongest tip in the deck, not the easiest one to illustrate.
+- [ ] Shuffling two middle tips breaks nothing.
+
+**The tips**
+
+- [ ] Every headline is one imperative sentence, ≤16 words, ≤3 lines, hand-broken.
+- [ ] Every tip is on rung 3 or 4 of the specificity ladder.
+- [ ] No headline is a question, and none contains a hedge.
+- [ ] Every subtext gives a REASON and does not restate its headline.
+- [ ] Every number on every slide is one you would defend if challenged.
+- [ ] No competitor is named.
+
+**The shill**
+
+- [ ] It sits in the MIDDLE: at least two real tips before it and two after it, cover excluded.
+- [ ] It is the only slide naming the offer.
+- [ ] With the brand name covered, it is still one of the two best tips in the deck.
+- [ ] It states a MECHANISM the reader could follow with no product at all.
+- [ ] The brand is lowercase, in the subtext, in the 500 weight, named once.
+- [ ] Slide 3 is visually indistinguishable from the others — same layout, type, colour, art class.
+
+**The pictures**
+
+- [ ] One art class, declared in the build log before sourcing.
+- [ ] Every cutout is the tip's noun. Covering the text, a stranger could roughly guess each slide.
+- [ ] 1–3 cutouts per slide, plus at most one hand-drawn mark.
+- [ ] At least one all-type slide, if any tip lacks a physical noun.
+- [ ] Every cutout's alpha is mostly 0-or-255. No grey halo on the white page.
+- [ ] Objects are rotated ±3–8°, none overlapping another's silhouette.
+- [ ] One light direction, or no contact shadows at all.
+- [ ] Every required credit is in the post caption. Nothing attributive is on a slide.
+
+**The page**
+
+- [ ] ONE background mode across the whole deck — never per slide.
+- [ ] Photo mode: every text band measured at ≥4.5:1, and you know which slide is the weakest.
+- [ ] Photo mode: every background is the PLACE its tip is about, not a mood.
+- [ ] Photo mode: hand-drawn marks use `{ink}`; photographic cutouts are never recoloured.
+- [ ] Photo mode: the blur is on the image and the dim is a separate flat layer (no `modal-scrim`).
+- [ ] Paper mode: one canvas colour across the deck. No footage, gradient, texture, border or card.
+- [ ] Headline band starts at 14% on every slide, whatever its line count.
+- [ ] Nothing in the top 14% or bottom 20%.
+- [ ] Two weights of ONE self-hosted family, 800 and 500, no fallback chain in the stack.
+- [ ] 2–4 emphasised runs per headline, weight-only, verb and noun.
+- [ ] No stroke, no text shadow, no plate behind any text.
+
+**The exports**
+
+- [ ] Frame 0 is slide 1, fully settled, headline legible — it is the thumbnail AND the cover.
+- [ ] No entrance transition on slide 1; no transition between any slides.
+- [ ] N PNGs exported at start+1.5s, each one complete on its own.
+- [ ] Slides uploaded to DURABLE My Files, ordered, recorded in `slides.json`.
+- [ ] The MP4 is `N × 3.0s` and the cuts land on the second.
+- [ ] The submission says the carousel is the deliverable and the MP4 is the preview.
+
+---
+
+## The whole-deck review — do this last, on the contact sheet
+
+Gates 1 and 2 grade slides. This grades the deck, and it is the only pass that catches what
+actually ships broken. **Open `slides/contact-sheet.png` and read it as one image.**
+
+- [ ] **Does the headline sit at the same height on all N slides?** The wobble is invisible one
+      slide at a time and obvious on the sheet.
+- [ ] **How many type sizes are there?** There should be two. Count them.
+- [ ] **Do the cutouts look like one set?** Same class, same rendering, same light, plausible
+      relative scale.
+- [ ] **Is the canvas the same white on all N?** A cutout with a near-white background rectangle
+      makes its slide read a different colour and only shows up beside its neighbours.
+- [ ] **Which slide is the weakest?** There is one. Either rewrite it or delete it — a five-slide
+      deck with no weak slide beats a six-slide deck with one.
+- [ ] **Can you tell which slide is the ad?** If your eye goes to slide 3 before you read it, it is
+      styled differently and Rule 1 is broken.
+- [ ] **Would you save this?** Not "is it correct". Would you save it.
+
+Then watch the MP4 end to end at 1× with sound off, which is how it will be seen.
+
+---
+
+## Diagnosing a flop
+
+Diagnose by slot, never by declaring the deck bad.
+
+| Symptom | The weak slot | The fix |
+|---|---|---|
+| Almost no impressions | **Slide 1 / the thumbnail** | Slide 1 is not your best tip, or its headline is unreadable at feed scale. Swap in the strongest one |
+| Views, drop-off at slide 2–3 | **The shill** | It reads as an ad. Cover the brand: if the slide is now weak, rewrite it as a tip |
+| Swiped to the end, no saves | **The tips** | They are on rung 1–2. Add a mechanism and a number to each |
+| Saves, no comments | **The caption** | The bait is missing or generic. It lives in the caption, so fix the caption, not the deck |
+| Good numbers, no clicks to the offer | **Nothing is broken** | This format seeds a mechanism; it does not close. Judge it on saves and branded search, not clicks |
+| "Looks AI-made" in the comments | **The sequence** | Almost always the band wobble, mixed art classes, or a stroke on the type. Contact sheet, gate 3 |
+
+---
+
+## Appendix A — `build-deck.py` (a deck from a manifest)
+
+Builds the whole composition from a TSV. The manifest is the artifact you iterate on; the HTML is
+generated and should never be hand-edited after the first run.
+
+```
+# slides.tsv — one line per slide, TAB separated. '#' comments ignored.
+# headline (use *asterisks* for the 800-weight runs, | for a hand line break)
+#   \t subtext (| for a line break, empty for none)
+#   \t cutouts: file@left%,top%,width%,rotate ; ... (empty for an all-type slide)
+*Never* eat on the block with the|*menus outside*.	If someone has to wave you in,|the food is not doing it.	media/cutouts/menu-board.png@34,38,32,-5
+```
+
+```python
+#!/usr/bin/env python3
+"""
+Build a sticker-slideshow-tips composition from a TSV manifest.
+
+Four things this exists to get right:
+
+  * LAYERS ARE FLAT. Every layer is a direct child of #root carrying its own
+    data-start / data-duration / data-layer-kind. A "slide" is not a container —
+    it is just the set of layers that share a start time. Wrapping a slide's
+    layers in an outer div.clip renders correctly and is INVISIBLE to the
+    toolchain: `vidfarm qa` reports scenes:0, and first_frame_visual /
+    first_frame_text then fail on a composition whose frame 0 is actually fine.
+  * THE BANDS DO NOT MOVE. Every headline is pinned to the same top, whatever
+    its line count. Vertically centring per slide is the "wobble" that only
+    shows up on the contact sheet (Rule 3).
+  * GEOMETRY IS INLINE, TYPOGRAPHY IS IN THE CLASS. `.clip{inset:0}` has to be
+    reset per layer with `inset:auto`, that reset lives in the inline style
+    attribute, and inline BEATS the stylesheet — so a `top:` in a CSS class
+    resolves to `top:auto` and the subtext silently stacks on the headline.
+  * THE BLUR AND THE DIM ARE DIFFERENT ELEMENTS. A single backdrop that is both
+    blurred and dimmed is the website-modal staging pattern (`vidfarm qa` →
+    modal-scrim). Blur the photo on the <img>; carry the dim on a separate flat
+    colour layer above it. Same look, and it is not modal staging.
+
+Background modes (--bg-mode):
+
+    paper           flat near-white canvas, dark text, NO shadow (the reference)
+    photo-darken    photo, blurred + darkened, WHITE text
+    photo-lighten   photo, blurred + lightened, DARK text
+    photo-gray      photo, blurred + desaturated under a neutral wash, WHITE text
+
+Any photo mode needs a 4th TSV column naming that slide's background image.
+
+usage: build-deck.py <slides.tsv> <out.html> [--title=…] [--canvas=#FDFCFA]
+                     [--bg-mode=paper|photo-darken|photo-lighten|photo-gray]
+                     [--head-px=70] [--sub-px=40]
+"""
+import html, sys
+
+SLIDE_SEC = 3.0
+W, H = 1080, 1920
+BANDS = dict(head_top=14.0, cut_top=34.0, cut_bot=66.0, sub_top=73.0)
+
+# Each mode is a complete legibility contract: what happens to the photo, what
+# flat wash sits over it, what colour the text is, and what that text needs to
+# survive the picture. They are tuned together — changing one value alone is how
+# a deck ends up with 2.9:1 headlines that "looked fine" on one slide.
+MODES = {
+    "paper": dict(
+        filter=None, wash=None, ink="#111111",
+        # On a flat page the contrast is already maximal and a shadow is the
+        # thing that makes type read as machine-made (Rule 5).
+        shadow="none", cut_shadow="none"),
+    "photo-darken": dict(
+        filter="blur(18px) brightness(0.38) saturate(0.90)",
+        wash="linear-gradient(180deg,rgba(0,0,0,0.62) 0%,rgba(0,0,0,0.62) 32%,rgba(0,0,0,0.22) 44%,rgba(0,0,0,0.22) 64%,rgba(0,0,0,0.62) 72%,rgba(0,0,0,0.62) 100%)", ink="#FFFFFF",
+        shadow="0 2px 18px rgba(0,0,0,0.55)",
+        cut_shadow="drop-shadow(0 14px 30px rgba(0,0,0,0.55))"),
+    "photo-lighten": dict(
+        filter="blur(18px) brightness(1.62) saturate(0.42)",
+        wash="linear-gradient(180deg,rgba(255,255,255,0.78) 0%,rgba(255,255,255,0.78) 32%,rgba(255,255,255,0.46) 44%,rgba(255,255,255,0.46) 64%,rgba(255,255,255,0.78) 72%,rgba(255,255,255,0.78) 100%)", ink="#111111",
+        shadow="0 2px 18px rgba(255,255,255,0.75)",
+        cut_shadow="drop-shadow(0 14px 30px rgba(0,0,0,0.30))"),
+    "photo-gray": dict(
+        filter="blur(14px) grayscale(1) brightness(0.48)",
+        wash="linear-gradient(180deg,rgba(18,18,20,0.64) 0%,rgba(18,18,20,0.64) 32%,rgba(18,18,20,0.18) 44%,rgba(18,18,20,0.18) 64%,rgba(18,18,20,0.64) 72%,rgba(18,18,20,0.64) 100%)", ink="#FFFFFF",
+        shadow="0 2px 18px rgba(0,0,0,0.50)",
+        cut_shadow="drop-shadow(0 14px 30px rgba(0,0,0,0.50))"),
+}
+
+args = [a for a in sys.argv[1:] if not a.startswith("--")]
+opts = dict(a.split("=", 1) for a in sys.argv[1:] if a.startswith("--") and "=" in a)
+manifest, out = args[0], args[1]
+canvas = opts.get("--canvas", "#FDFCFA")
+title = opts.get("--title", "tip deck")
+head_px = int(opts.get("--head-px", 70))
+sub_px = int(opts.get("--sub-px", 40))
+bg_mode = opts.get("--bg-mode", "paper")
+if bg_mode not in MODES:
+    sys.exit(f"unknown --bg-mode {bg_mode!r}; pick one of {', '.join(MODES)}")
+M = MODES[bg_mode]
+
+rows = []
+for raw in open(manifest):
+    raw = raw.rstrip("\n")
+    if not raw.strip() or raw.lstrip().startswith("#"):
+        continue
+    p = raw.split("\t")
+    rows.append(tuple((p[i].strip() if len(p) > i else "") for i in range(4)))
+
+if not 5 <= len(rows) <= 9:
+    print(f"⚠ {len(rows)} slides — the harness range is 5-9", file=sys.stderr)
+if M["filter"] and any(not r[3] for r in rows):
+    missing = [i + 1 for i, r in enumerate(rows) if not r[3]]
+    sys.exit(f"--bg-mode {bg_mode} needs a background image per slide; "
+             f"column 4 is empty on slide(s) {missing}")
+
+
+def rich(s):
+    """*bold* -> 800 weight; | -> a hand line break. Escape everything else."""
+    s = html.escape(s)
+    return "".join(f"<b>{c}</b>" if i % 2 else c
+                   for i, c in enumerate(s.split("*"))).replace("|", "<br>")
+
+
+CSS = f"""
+ /* Self-hosted, NO fallback chain. A font that fails to fetch would silently
+    substitute and the deck would quietly stop looking like itself. Two weights
+    of one family is the entire regime (Typography DNA). */
+ @font-face{{font-family:'Deck';font-weight:800;font-display:block;
+   src:url('media/fonts/deck-800.woff2') format('woff2')}}
+ @font-face{{font-family:'Deck';font-weight:500;font-display:block;
+   src:url('media/fonts/deck-500.woff2') format('woff2')}}
+ body{{margin:0;background:{canvas}}}
+ #root{{position:relative;width:{W}px;height:{H}px;overflow:hidden;background:{canvas}}}
+ .clip{{position:absolute;inset:0}}
+
+ /* THE BANDS are set INLINE, not here — see the module docstring. Geometry
+    inline, typography in the class. */
+ .head{{font-family:'Deck';font-weight:500;font-size:{head_px}px;line-height:1.22;
+   letter-spacing:-0.01em;color:{M['ink']};text-align:center;
+   text-shadow:{M['shadow']}}}
+ .head b{{font-weight:800}}
+ .sub{{font-family:'Deck';font-weight:500;font-size:{sub_px}px;line-height:1.34;
+   color:{M['ink']};text-align:center;text-shadow:{M['shadow']}}}
+ .cut{{object-fit:contain;filter:{M['cut_shadow']}}}
+
+
+ /* The photo is scaled past the frame so the blur's own soft edge is cropped
+    off — a blurred image at exactly 100% shows a pale vignette on all four
+    sides, which reads as a rendering bug rather than a background. */
+ .bg{{width:110%;height:110%;left:-5%;top:-5%;object-fit:cover;filter:{M['filter'] or 'none'}}}
+ .wash{{background:{M['wash'] or 'transparent'}}}
+"""
+
+
+def layer(kind, idx, start, cls, style, inner="", tag="div", extra="", label=""):
+    """One FLAT layer: a direct child of #root, never nested inside another clip."""
+    common = (f'class="clip {cls}" data-hf-id="{idx}" data-layer-mode="publish" '
+              f'data-layer-kind="{kind}" data-start="{start}" '
+              f'data-duration="{SLIDE_SEC}" data-end="{round(start+SLIDE_SEC,3)}" '
+              f'data-label="{html.escape(label)}"')
+    if tag == "img":
+        return f'  <img id="{idx}" {common} {extra} style="{style}">'
+    return f'  <div id="{idx}" {common} style="{style}">{inner}</div>'
+
+
+layers = []
+for i, (head, sub, cuts, bg) in enumerate(rows):
+    t = round(i * SLIDE_SEC, 3)
+    n = i + 1
+    if M["filter"] and bg:
+        layers.append(layer("image", f"s{n:02d}-bg", t, "bg",
+                            "inset:auto;z-index:0", tag="img",
+                            extra=f'src="{html.escape(bg)}"',
+                            label=f"slide {n} background"))
+        if M["wash"]:
+            layers.append(layer("shape", f"s{n:02d}-wash", t, "wash",
+                                "inset:0;z-index:1", label=f"slide {n} wash"))
+    layers.append(layer("text", f"s{n:02d}-head", t, "head",
+                        f"inset:auto;left:8%;width:84%;top:{BANDS['head_top']}%;z-index:3",
+                        rich(head), label=f"slide {n} headline"))
+    if sub:
+        layers.append(layer("text", f"s{n:02d}-sub", t, "sub",
+                            f"inset:auto;left:12%;width:76%;top:{BANDS['sub_top']}%;z-index:3",
+                            rich(sub), label=f"slide {n} subtext"))
+    for j, spec in enumerate(filter(None, (c.strip() for c in cuts.split(";")))):
+        src, geo = spec.split("@")
+        src = src.replace("{ink}", "light" if M["ink"] == "#FFFFFF" else "dark")
+        left, top, width, rot = (geo.split(",") + ["0"])[:4]
+        if not BANDS["cut_top"] - 2 <= float(top) <= BANDS["cut_bot"]:
+            print(f"⚠ slide {n}: cutout top {top}% is outside the "
+                  f"{BANDS['cut_top']}-{BANDS['cut_bot']}% band", file=sys.stderr)
+        layers.append(layer(
+            "image", f"s{n:02d}-cut{j+1}", t, "cut",
+            f"inset:auto;left:{left}%;top:{top}%;width:{width}%;height:auto;"
+            f"transform:rotate({rot}deg);z-index:2",
+            tag="img", extra=f'src="{html.escape(src)}"',
+            label=f"slide {n} cutout {j+1}"))
+
+total = round(len(rows) * SLIDE_SEC, 3)
+open(out, "w").write(f"""<!doctype html>
+<html lang="en"><head><meta charset="UTF-8">
+<meta name="viewport" content="width={W}, height={H}">
+<title>{html.escape(title)}</title>
+<style>{CSS}</style></head><body>
+<div id="root" data-composition-id="tip_deck" data-start="0"
+     data-width="{W}" data-height="{H}" data-duration="{total}">
+{chr(10).join(layers)}
+</div></body></html>
+""")
+print(f"{len(rows)} slides · {total}s · {len(layers)} flat layers · "
+      f"bg-mode {bg_mode} · ink {M['ink']} · {out}")
+print("stills at: " + ",".join(str(round(i * SLIDE_SEC + 1.5, 1)) for i in range(len(rows))))
+```
+
+## Appendix B — `export-deck.sh` (both deliverables from one composition)
+
+```bash
+#!/usr/bin/env bash
+# Export the carousel (the deliverable) and the MP4 (the preview) from one composition.
+# usage: export-deck.sh <work-dir> <n-slides> <folder-slug>
+set -euo pipefail
+DIR="$1"; N="$2"; SLUG="$3"
+
+# Stills at start+1.5s of each 3.0s slide — the SETTLED state, never the entrance.
+AT=$(python3 -c "print(','.join(str(i*3.0+1.5) for i in range($N)))")
+
+vidfarm stills "$DIR" --at "$AT" --out "$DIR/slides" --sheet
+echo "▶ OPEN $DIR/slides/contact-sheet.png AND READ IT AS ONE IMAGE — this is gate 3."
+
+# Rename to an explicit order. The order is DATA: slide 3 being the shill is load-bearing.
+i=1; for f in "$DIR"/slides/*.png; do
+  case "$f" in *contact-sheet*) continue;; esac
+  mv "$f" "$DIR/slides/slide-$(printf %02d $i).png"; i=$((i+1))
+done
+
+# DURABLE My Files — the temp store deletes after 30 days and the carousel would 404.
+: > "$DIR/slides.json"; echo "[" >> "$DIR/slides.json"
+for f in "$DIR"/slides/slide-*.png; do
+  URL=$(vidfarm put-file "$f" --folder "decks/$SLUG" \
+         --notes "$SLUG tip deck — $(basename "$f")" --json | python3 -c \
+         'import json,sys; print(json.load(sys.stdin).get("viewUrl",""))')
+  echo "  \"$URL\"," >> "$DIR/slides.json"
+done
+sed -i '' '$ s/,$//' "$DIR/slides.json"; echo "]" >> "$DIR/slides.json"
+
+# The preview copy. Same composition, no second build.
+vidfarm render "$DIR" --wait
+
+echo "carousel: $DIR/slides.json  ($N ordered durable URLs)"
+echo "preview : $DIR/final.mp4"
+```
+
+## Appendix C — `slides-qa.py` (gate 2: enforce it on the EXPORTED PNGs)
+
+`vidfarm qa` reads the DOM and cannot see a grey halo, a drifting canvas or a wobbling band. This
+runs on the pixels that ship.
+
+```bash
+python3 slides-qa.py ./work/slides --brand dishcover.io --shill-slide 3
+```
+
+```python
+#!/usr/bin/env python3
+"""
+Grade an exported tip deck on the PNGs, not on the markup.
+
+Six checks, each one a defect that passes every DOM check and every per-slide
+eyeball, and shows up only across the set:
+
+  1. no slide is blank            — an entrance that had not settled at +1.5s
+  2. the canvas is ONE colour     — a cutout with a near-white background plate
+                                    makes its slide read a different white, and
+                                    that is only visible beside its neighbours
+  3. the headline band is level   — the vertical wobble of Rule 3
+  4. nothing in the reserved zones— top 14% / bottom 20% belong to the platform
+  5. exactly ONE slide names the  — Rule 1, checked against the deck's text
+     brand                          manifest rather than by OCR
+  6. every cutout's alpha is
+     mostly 0-or-255              — run separately on media/cutouts/*.png
+
+usage: slides-qa.py <slides-dir> [--brand <token>] [--shill-slide <n>] [--ink light|dark]
+                                 [--manifest <slides.tsv>] [--cutouts <dir>]
+"""
+import glob, os, subprocess, sys
+
+d = sys.argv[1]
+opt = dict(zip(sys.argv[2::2], sys.argv[3::2]))
+brand = opt.get("--brand")
+shill = int(opt.get("--shill-slide", 3))
+manifest = opt.get("--manifest")
+cutdir = opt.get("--cutouts")
+ink = opt.get("--ink")   # "light" | "dark" — the mode's text colour
+
+fails, warns = [], []
+
+
+def probe(png, crop):
+    """Mean RGB of a crop, via ffmpeg — no PIL dependency."""
+    raw = subprocess.run(
+        ["ffmpeg", "-v", "error", "-i", png, "-vf", f"crop={crop},scale=1:1",
+         "-frames:v", "1", "-f", "rawvideo", "-pix_fmt", "rgb24", "-"],
+        capture_output=True).stdout
+    return tuple(raw[:3]) if len(raw) >= 3 else (0, 0, 0)
+
+
+def rowdark(png, y, h, w=1080):
+    """Fraction of ink in a horizontal strip — how we find the headline."""
+    raw = subprocess.run(
+        ["ffmpeg", "-v", "error", "-i", png, "-vf",
+         f"crop={w}:{h}:0:{y},format=gray,scale=64:1", "-frames:v", "1",
+         "-f", "rawvideo", "-pix_fmt", "gray", "-"],
+        capture_output=True).stdout
+    return sum(1 for b in raw if b < 140) / max(len(raw), 1)
+
+
+slides = sorted(f for f in glob.glob(os.path.join(d, "*.png"))
+                if "contact-sheet" not in f)
+if not slides:
+    sys.exit("no slides found")
+if not 5 <= len(slides) <= 9:
+    warns.append(f"{len(slides)} slides — the harness range is 5-9")
+
+# ---------------------------------------------- 1 + 2. blank slides, canvas drift
+canvases = []
+for s in slides:
+    c = probe(s, "40:40:20:20")          # a corner, always canvas
+    canvases.append(c)
+    # a slide whose whole frame equals its corner has nothing on it
+    mid = probe(s, "600:600:240:520")
+    if max(abs(a - b) for a, b in zip(c, mid)) < 4:
+        fails.append(f"{os.path.basename(s)}: BLANK — nothing on the slide at +1.5s")
+base = canvases[0]
+# PAPER MODE ONLY. A photo deck is SUPPOSED to change picture every slide, and
+# the same goes for the two checks below: the headline detector finds photo
+# detail rather than glyphs, and "ink in the reserved zone" is true of every
+# full-bleed photograph. On a photo deck the contrast gate replaces all three.
+if not ink:
+    for s, c in zip(slides, canvases):
+        if max(abs(a - b) for a, b in zip(base, c)) > 3:
+            fails.append(f"{os.path.basename(s)}: canvas {c} drifts from {base}")
+
+# --------------------------------------------------- 3. the headline band is level
+tops = []
+for s in (slides if not ink else []):
+    top = None
+    for y in range(200, 700, 12):        # scan the headline band, 10%-36%
+        if rowdark(s, y, 12) > 0.06:
+            top = y
+            break
+    tops.append(top)
+    if top is None:
+        fails.append(f"{os.path.basename(s)}: no headline found in the band")
+known = [t for t in tops if t is not None]
+if known and max(known) - min(known) > 24:   # >24px = ~1.25% of canvas height
+    warns.append(f"headline band WOBBLES: tops span {min(known)}-{max(known)}px. "
+                 f"Pin every headline to the same top (Rule 3)")
+
+# ------------------------------------------------------- 4. the reserved zones
+for s in (slides if not ink else []):
+    if rowdark(s, 0, 240) > 0.02:
+        warns.append(f"{os.path.basename(s)}: ink in the top 14% (platform chrome)")
+    if rowdark(s, 1536, 384) > 0.02:
+        warns.append(f"{os.path.basename(s)}: ink in the bottom 20% (caption/buttons)")
+
+# ------------------------------------------------------------ 5. exactly ONE shill
+if brand and manifest:
+    hits = [i + 1 for i, line in enumerate(
+        l for l in open(manifest) if l.strip() and not l.lstrip().startswith("#"))
+        if brand.lower() in line.lower()]
+    if hits != [shill]:
+        fails.append(f"brand '{brand}' appears on slides {hits or 'none'} — "
+                     f"Rule 1 requires exactly [{shill}]")
+
+# ------------------------------------------------- 5b. TEXT CONTRAST (readability)
+# The one thing a photo background can silently destroy. Measuring beats looking:
+# a headline that reads fine on slide 2's dark doorway can vanish on slide 5's
+# bright tiled wall, and you only notice on the slide you did not open.
+#
+# The background is recovered by BLURRING the band hard enough to erase the
+# glyphs — a boxblur wider than the stroke leaves the backdrop and nothing else —
+# then the WORST patch of that backdrop is compared against the known ink colour.
+# Worst case for light ink is the brightest patch; for dark ink, the darkest.
+def _lin(v):
+    c = v / 255.0
+    return c / 12.92 if c <= 0.03928 else ((c + 0.055) / 1.055) ** 2.4
+
+
+def band_contrast(png, y, h, ink_is_light):
+    raw = subprocess.run(
+        ["ffmpeg", "-v", "error", "-i", png, "-vf",
+         f"crop=1080:{h}:0:{y},format=gray,boxblur=24:2,scale=48:8",
+         "-frames:v", "1", "-f", "rawvideo", "-pix_fmt", "gray", "-"],
+        capture_output=True).stdout
+    if not raw:
+        return None
+    vals = sorted(_lin(b) for b in raw)
+    L_ink = 1.0 if ink_is_light else _lin(0x11)
+    # 90th / 10th percentile rather than the absolute extreme: one stray blurred
+    # highlight is not where the text sits, and grading against it fails a deck
+    # that is genuinely fine.
+    L_bg = vals[int(len(vals) * 0.90)] if ink_is_light else vals[int(len(vals) * 0.10)]
+    hi, lo = max(L_ink, L_bg), min(L_ink, L_bg)
+    return (hi + 0.05) / (lo + 0.05)
+
+
+if ink:
+    light = ink == "light"
+    for s in slides:
+        for name, y, h in (("headline", 260, 340), ("subtext", 1395, 175)):
+            r = band_contrast(s, y, h, light)
+            if r is None:
+                continue
+            if r < 4.5:
+                fails.append(f"{os.path.basename(s)}: {name} contrast {r:.1f}:1 — "
+                             f"unreadable (WCAG AA needs 4.5:1). Deepen the wash "
+                             f"or switch polarity")
+            elif r < 7.0:
+                warns.append(f"{os.path.basename(s)}: {name} contrast {r:.1f}:1 — "
+                             f"passes AA but thin for arm's-length reading")
+
+# ------------------------------------------------------------ 6. cutout alpha health
+if cutdir:
+    for png in sorted(glob.glob(os.path.join(cutdir, "*.png"))):
+        raw = subprocess.run(
+            ["ffmpeg", "-v", "error", "-i", png, "-vf",
+             "format=rgba,alphaextract,scale=64:64", "-frames:v", "1",
+             "-f", "rawvideo", "-pix_fmt", "gray", "-"],
+            capture_output=True).stdout
+        if not raw:
+            continue
+        partial = sum(1 for b in raw if 8 < b < 247) / len(raw)
+        if partial > 0.08:
+            fails.append(f"{os.path.basename(png)}: {partial:.0%} partial alpha — "
+                         f"this WILL show as a grey halo on the white page")
+
+# ------------------------------------------------------------------------ verdict
+for f in fails:
+    print(f"  ✗ {f}")
+for w in warns:
+    print(f"  ⚠ {w}")
+print(f"\n{len(slides)} slides · {len(fails)} fail · {len(warns)} warn")
+print("\n▶ NOW OPEN contact-sheet.png. This checked pixels, not whether it reads as one deck.")
+sys.exit(1 if fails else 0)
+```
+
+## Appendix D — a worked example: dishcover.io, NYC eats
+
+Six slides, 18.0s — the reference template's exact length. Cost mode `minimize`, interaction mode
+`interactive`. Art class: **photographic cutout**, one generated sheet.
+
+**Build log header**
+
+```
+TOPIC:     Finding the best eats in NYC
+READER:    Lives here, eats out twice a week, keeps ending up somewhere fine
+OFFER:     dishcover.io — search restaurants by DISH instead of by restaurant
+ART CLASS: photographic cutout, one sheet, contact shadows off
+SHILL:     slide 3
+```
+
+**Post caption** (this is where the topic, the bait and the credits live)
+
+```
+6 rules for eating well in NYC 🗽
+which one did you already know?
+```
+
+**`slides.tsv`**
+
+```
+*Never* eat on the block|with the *menus outside*.	If someone has to wave you in,|the food is not doing the work.	media/cutouts/menu-board.png@34,38,32,-5
+Go at *5:30* or go at *9:45*.	The kitchen is calmest at the edges|of service. Same dish, better plate.	media/cutouts/clock.png@22,40,24,-4 ; media/cutouts/plated-dish.png@54,38,28,6
+*Search the dish*, not|the *restaurant*.	"Best birria in Queens" finds you dinner.|"Best Mexican near me" finds you a parking lot.|dishcover.io is built to search that way.	media/cutouts/query-scribble.png@14,42,26,-3 ; media/cutouts/arrow.png@42,46,14,0 ; media/cutouts/birria.png@58,38,28,5
+*One dish* decides the place.	Every good spot does one thing|better than anyone. Order that.	media/cutouts/signature-dish.png@36,37,30,-6
+Trust a *30-item* menu|over a *200-item* one.	A short menu means a kitchen that repeats.|A long one means a freezer.	media/cutouts/menu-short.png@20,38,26,-5 ; media/cutouts/menu-long.png@54,36,26,7
+*Walk two more blocks.*	Every train exit has a rent-funded|trap in the first 200 feet.	media/cutouts/subway-exit.png@34,39,32,-4
+```
+
+**Slide 3, read against the lock**
+
+- *Actionable?* Yes — search a dish name in any search box tonight.
+- *Shill-as-tip?* Delete `dishcover.io is built to search that way.` The remaining two lines are
+  arguably the best tip in the deck. Passes.
+- *Still-complete?* Scribble → arrow → birria bowl. The mechanism is legible with the text covered.
+
+**Build and export**
+
+```bash
+python3 build-deck.py slides.tsv work/index.html --title="dishcover — NYC eats deck" \
+                                                 --canvas="#FDFCFA"
+vidfarm qa ./work --harness ./experimental/sticker-slideshow-tips.md
+bash export-deck.sh ./work 6 nyc-eats
+python3 slides-qa.py ./work/slides --brand dishcover.io --shill-slide 3 \
+        --manifest slides.tsv --cutouts ./work/media/cutouts
+open ./work/slides/contact-sheet.png     # gate 3 — the pass that actually matters
+```
+
+**What the first pass got wrong** (recorded because it is what this format gets wrong):
+
+1. Slide 2 had two cutouts at the same scale, so the clock read as a plate. Scaled the clock down
+   to 24% and the dish up to 28%.
+2. Slide 5's two menus came from different sources — one photographic, one a flat vector. Mixed art
+   class, invisible on slide 5, obvious on the contact sheet. Regenerated both from the sheet.
+3. The headline on slide 6 was one line where every other slide was two, so its cutout band started
+   40px high. Fixed by pinning the band, not by padding the headline.
+4. `dishcover.io` was originally in the 800 weight. It read as a banner. Dropped to 500 with the
+   rest of the subtext and the slide stopped looking like the ad slide.
+
+## Appendix E — the harness in one paragraph, for a handoff
+
+> Build a 6-slide tip carousel, 9:16, 1080×1920, on a flat `#FDFCFA` page — no footage, no
+> gradients, no cards. Each slide: one imperative tip as a centred headline (TikTok Sans or
+> Montserrat, two weights 800/500 mixed inside the sentence, 76px, pinned at 14% from the top,
+> hand-broken lines) plus an optional reason line at 73% (40px, 500) and 1–3 die-cut photographic
+> cutouts in the 34–66% band, rotated ±3–8°. No stroke, no shadow, no plate on any text. Slide 3
+> names dishcover.io once, lowercase, in the subtext, and is written so it is still a good tip with
+> the brand removed. No title slide; the topic and the comment bait are the post caption. Ship the
+> six PNGs as the deliverable (durable My Files, ordered in `slides.json`) and an MP4 at exactly
+> 3.0s per slide with hard cuts as the preview. Full rules:
+> `vidfarm.cc/experimental/sticker-slideshow-tips.md`.