@kolbo/mcp 1.81.6 → 1.81.8

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@kolbo/mcp",
-  "version": "1.81.6",
+  "version": "1.81.8",
   "description": "Kolbo AI MCP Server - Generate images, videos, music, speech, and sound effects from Claude Code",
   "main": "src/index.js",
   "bin": {

package/skill/GENERATED.md

@@ -1,6 +1,6 @@
 # AUTO-GENERATED — do not edit
 
-This tree is mirrored from kolbo-code@a463d99, the single source of truth.
+This tree is mirrored from kolbo-code@beae0e9, the single source of truth.
 Canonical source: packages/opencode/skills/kolbo/
 Distribution: .github/workflows/sync-skill-to-plugin.yml
 

package/skill/SKILL.md

@@ -1,5 +1,5 @@
 ---
-version: 0.9.7
+version: 0.9.10
 name: kolbo
 description: |
   Generate, edit, analyze, and direct creative media through Kolbo AI: images,
@@ -300,14 +300,6 @@ Four surfaces show the same job. Use this map — never invent a fifth:
 
 **Checking status — NEVER poll in a loop.** `get_generation_status` takes `wait=true` (blocks server-side until done, ~3 min) and `generation_ids` (check MANY generations in ONE call — returns `all_done` + which are still running). One `wait=true` call replaces any polling loop: check ALL in-flight ids in ONE call, never one by one, never without `wait`. If it comes back with some still processing, call it ONCE more with `wait=true` and the remaining ids.
 
-**🛑 Runaway-loop guard — ONE generation per requested item.** When the user asks for **one specific change**, the answer is **a single tool call**. After URLs return, **stop**. Surface and wait. You are NOT allowed to:
-- Fire the same tool 3+ times in a single turn unless the user explicitly asked for "N variations".
-- Re-fire because you think the result might not be exactly what the user wanted.
-- Auto-retry on success.
-- Fire 5+ parallel `generate_video*` calls speculatively.
-
-**Only re-fire when:** user explicitly asked for variations with a count, OR previous call returned `failure.retryable === true` (ONE retry), OR previous call returned `completed` but `urls.length === 0` (ONE retry).
-
 **Detecting failure — a generation can fail three ways. Treat ALL as failure:**
 
 1. **Tool returns `error`** — explicit. Surface, suggest retry, log `generation_id`.
@@ -368,6 +360,18 @@ Write each session's `session_id` + plan name into `.kolbo/production.md` `### S
 
 **Parameter gotcha:** `num_images` (1–4, same prompt different seeds) on `generate_image` vs `scene_count` (1–8, distinct prompt per scene) on `generate_creative_director`. **Never pass `num_images` to Creative Director.**
 
+## 🛑 Runaway-Loop Guard — ONE Generation per Requested Item (CRITICAL)
+
+When the user asks for **one specific change**, the answer is **a single tool call**. After URLs return, **stop**. Surface and wait.
+
+You are NOT allowed to:
+- Fire the same tool 3+ times in a single turn unless the user explicitly asked for "N variations".
+- Re-fire because you think the result might not be exactly what the user wanted.
+- Auto-retry on success.
+- Fire 5+ parallel `generate_video*` calls speculatively.
+
+**Only re-fire when:** user explicitly asked for variations with a count, OR previous call returned `failure.retryable === true` (ONE retry), OR previous call returned `completed` but `urls.length === 0` (ONE retry).
+
 ## ⚠️ Editing an Existing Video → ONE Call, Not Frames-First (CRITICAL)
 
 Existing video → modify → **single `generate_video_from_video` call** with source video URL + edit prompt.

package/skill/VERSION

@@ -1 +1 @@
-0.9.7
+0.9.10

package/skill/references/workflows/thumbnails.md

@@ -5,113 +5,297 @@ channel art**: "thumbnail for my video", "YouTube cover", "Shorts cover", "make
 clickable", "higher CTR".
 
 A thumbnail is not a nice image at small size. It is a different craft with a different
-success test: **it is judged at ~200px inside a scrolling feed, next to a hundred others.**
+success test: **it is judged at ~168px inside a scrolling feed, next to a hundred others.**
 Everything below follows from that.
 
 ---
 
-## The four layers
-
-Every thumbnail that works has exactly these, in this order of importance.
-
-### 1. ONE hero subject — 40–70% of the frame
+## ⚠️ Two things to get right before advising anyone
 
-A face with a big, readable emotion (eyes visible, mouth doing something), or a single
-object caught mid-action. Crop tight — chest-up for a person.
+**1. The metric is watch time, not clicks.** YouTube's own A/B thumbnail test optimises
+"for overall watch time over other metrics, like click-through rate." A thumbnail that
+wins the click and loses the viewer *loses the test*. This is the mechanical reason the
+shock-face era ended. Never promise something the video does not deliver — it is both a
+spam-policy violation ("malicious clickbait" names thumbnails explicitly) and a test loss.
 
-Kill on sight: full-body wide shots, three competing focal points, a floating UI panel, an
-abstract gradient, "a person at a desk". If the concept has no surprise, wit, or
-impossibility in it, it will not stop a scroll — write the concept as one sentence first,
-and if that sentence is boring, invent harder before generating.
+**2. Most thumbnail statistics online are fabricated.** This topic is saturated with
+AI-generated SEO spam inventing authoritative-looking numbers ("47.3% of creators…",
+"9% vs 4% CTR study…", "70% higher CTR for dark thumbnails"). Traced individually, they
+have no source. **Do not quote a thumbnail statistic to a user unless it is in this file.**
+Everything below is graded: **[A]** real study/primary source · **[A?]** real study, but
+its page could not be opened directly — figures corroborated only via secondary quotes,
+so treat as directional, not exact · **[B]** credible practitioner claim · **[C]** craft
+convention, no data.
 
-### 2. Extreme separation
+Also useful to know: CTR *falling* as a video spreads is normal, not failure — early
+impressions come from subscribers, then Browse/Suggested reach cold viewers. Half of all
+channels sit between **2% and 10% CTR**. **[A]**
 
-The subject must pop off the background: a dark field behind a bright subject, or one
-saturated accent against its complement. **Mid-tone on mid-tone is the number-one
-unreadable-thumbnail failure.** Muted, tasteful palettes lose in a feed.
-
-### 3. Text — 2 to 4 words, maximum
-
-- Quote the exact words in the prompt: `render only this text: "STOP DOING THIS"`.
-- Topmost layer, above every element and effect.
-- Roughly **35–45% of the canvas width**. Heavy condensed sans.
-- White or a single accent colour with a **thick dark outline or a solid backing bar** —
-  raw text on a busy image is illegible small.
-- Upper or lower third. **Never across the face.**
-- Forbid everything else explicitly: `no other text, no taglines, no watermark, no logo,
-  no captions, no placeholder text`.
+---
 
-Long strings come back mangled from every image model. If the user wrote a sentence, put
-three words on the image and tell them the rest belongs in the video title. For brand
-names, spell them letter-by-letter in the prompt and raise quality when the type is small.
+## The four layers
 
-**Depth trick:** let the subject overlap one word slightly (subject in front of one
-letter). Instant production value.
+### 1. ONE hero subject
+
+A face with one **legible** emotion, or a single object caught mid-action. Crop tight.
+
+**Faces scale with performance.** In a study of 500 breakout videos: 69% used a face,
+rising to 75% of the top 100 and **80% of the top 50**. Face occupies about **one third
+of the frame** in that top cohort. **[A?]** — the direction (faces matter, and matter more
+at the top) is safe; the exact percentages are not verified at source.
+
+**The exaggerated shocked face is dead as a default.** Only ~5% of those breakouts used
+an exaggerated expression; 6% of the top 50. MrBeast's team A/B tested ~30 videos
+open-mouth vs closed-mouth — **every closed-mouth version produced higher watch time**,
+and they changed the channel's house style permanently. **[A]**
+
+So: **a clear face beats a wild face.** Direct the expression to *one* readable emotion —
+closed-mouth determination, a genuine restrained smile, focused concentration, real
+scepticism. Direct eye contact into the lens, unless the gaze is deliberately pointing at
+the subject. Forbid in the prompt: gaping "O" mouth, bulging eyes, strained forehead,
+raised chin, any look of discomfort or effort.
+
+**Legibility test:** at 168×94, can someone name the emotion in under a second?
+
+**Faces are not mandatory** — 31% of breakouts had none. If the subject is a tool, an
+interface, or a generated frame, the object can be the hero.
+
+Kill on sight: three competing focal points, a floating UI panel over an abstract
+gradient, "a person at a desk". If the concept has no surprise, wit, or impossibility in
+it, invent harder before generating.
+
+### 2. Separation — luminance first, hue second
+
+**Greyscale test: desaturate the thumbnail. If the subject stops separating from the
+background, no palette will save it.** Complementary pairs (orange/teal, blue/orange)
+work mostly because they *also* carry a luminance gap.
+
+- **Three dominant colours maximum, text included.** **[B]**
+- **Restrain saturation.** The over-saturated HDR/clarity look — "exaggerated vibrancy,
+  bright lighting and sharp contrast" — is now documented as the visual signature of
+  AI-generated imagery, inherited from advertising imagery in training data. **[A]**
+  Push *luminance* contrast hard; keep saturation on one or two accent elements only.
+- **Check on both YouTube themes.** The UI is near-white or near-black with red accents.
+  A dark vignette that separates your subject in dark mode can vanish in light mode.
+  Saturated red as a dominant field reads as UI chrome, not content.
+- A single-hue thumbnail is *not* automatically weak — Kurzgesagt and MKBHD ship
+  near-monochrome successfully. What fails is a single hue with no luminance separation.
+
+### 3. Text — often none at all
+
+The real distribution among breakout thumbnails: **28% had no text** (the largest single
+group), **24% had 1–3 words**, median among those that had any was 5. **[A?]** Over half
+carried three words or fewer. The ubiquitous "3–5 bold words" advice describes a minority.
+
+- **Drop text entirely** when the image already states the promise and the title carries
+  the specifics. Text is a crutch for a thumbnail that hasn't found its image.
+- **Three words maximum** when used. Never repeat the title — the title is right next to
+  it. Text should add the *second half* of an idea the image starts.
+- Quote the exact words in the prompt: `render only this text: "…"`, and forbid the rest:
+  `no other text, no taglines, no watermark, no logo, no captions, no gibberish letters`.
+- Heavy-weight sans, one typeface. Cap height ≈ **12–15% of frame height**.
+- **Always** an outline, drop shadow, or solid colour block behind it.
+- Placement: top third or the side opposite the face. **Never the bottom-right quadrant.**
+- Let the subject overlap one letter slightly — instant depth.
+
+Long strings come back mangled from every image model. If words garble twice, generate the
+plate **text-free** and set type in the Canvas tool or in HTML — a clean plate plus real
+type beats a third mangled attempt. Logos are the same: **image models cannot render a
+real brand mark.** Generate the plate without it and composite the actual logo file.
 
 ### 4. Platform-safe composition
 
-| Format | Rules |
+| Zone | Rule |
 |---|---|
-| **16:9 — YouTube** | Centre-weighted. Nothing critical in the outer 15%: the duration chip sits bottom-right and the red progress bar covers the bottom edge on watched videos. |
-| **9:16 — Shorts / Reels / TikTok** | Every critical element must survive a **centre-square crop** (feeds and grids crop vertical media to its middle). Keep the top ~15% and the bottom ~20% clear of the platform's title, avatar, caption and buttons. |
-| **1:1** | Community posts and square feeds. |
+| **Bottom-right corner** | Duration chip sits here. Nothing important. |
+| **Bottom edge, full width** | Red progress bar on partially-watched videos. Keep critical content ≥40px off the bottom (at 1280×720 scale). |
+| **Overall safe area** | Everything load-bearing inside the centre ~1100×620 of a 1280×720 canvas. |
+| **9:16 — Shorts/Reels/TikTok** | Must survive a **centre-square crop** (grids and cross-posts crop vertical media to the middle). Keep the top ~15% and bottom ~20% clear of platform UI. |
+| **Validate at** | **168×94 px** — the desktop suggested-videos render size. Not 1280×720. |
 
-**Verification is a release gate, not a suggestion:** view the render at ~200px, and for
-9:16 crop the centre square first. If the words or the face don't survive, iterate.
+**File spec (updated — most guides are stale):** YouTube now recommends **3840×2160**,
+minimum width 640px, file cap **50 MB** on desktop. Driven by TV overtaking mobile as the
+primary US YouTube device in Feb 2025. **Export at 4K; design for 168px.** **[A]**
 
 ---
 
 ## Model choice
 
-Text fidelity is the entire constraint. Pick the image model that renders type most
-reliably, and raise the quality setting when the words are small or multi-font. If the
-words come back garbled twice, generate the image **text-free** and tell the user to set
-the type in the Canvas tool — a clean plate plus real type beats a third mangled attempt.
+Text fidelity is the constraint. `gpt-image-2` is the strongest for type — use
+**quality: medium** (its sweet spot) or high, and it handles **Hebrew** notably better
+than the alternatives. `nano-banana-2` is strong on cinematic people but weak on text.
+For a text-free plate, either is fine and the cheaper one wins.
+
+**Field result (Aug 2026, head-to-head, same prompt/refs/1K):** `nano-banana-2` beat
+`gpt-image-2` on BOTH photorealism (looked like a real studio photo, not a render) and
+reference-mark fidelity — and rendered a short Hebrew headline + Latin badge cleanly.
+The "weak on text" caveat applies to long/dense copy, not a 3-word headline. When a
+thumbnail is a photographic person + short text + referenced logos, run both models
+once and pick — do not assume gpt-image-2 wins by default.
+
+Lock the host with a **character Visual DNA** (`workflows/visual-dna.md`) so every
+thumbnail in a series is the same person — and check you are using the DNA of the person
+*as themselves*, not a costumed character DNA built for a film shoot.
+
+## Kolbo brand-asset kit — reuse these, do not rediscover them
+
+Learned the hard way (Time Machine tutorial thumbnail, Aug 2026): re-deriving logo
+files and the right Visual DNA from scratch every session, then patching AI mis-draws
+with HTML overlays, is slow and produces a worse result than just generating natively
+with the right references from the start. Check this list before generating anything
+with a Kolbo or model logo, or with Zohar's likeness, in it.
+
+| Asset | Where | Status |
+|---|---|---|
+| Kolbo K icon — **THE reference to use** | `Graphics\Logo\kolbo-ai-new-icon-black2.jpg` (white K on black) | ✅ verified real mark. **Always pass THIS file as the generation reference**, never a transparent cutout: the black background gives the model contrast to lock onto and it reproduces the K correctly. Proven tricks that keep the geometry exact fully in-model: (a) print it as a white chest logo on a black t-shirt, (b) render it inside a small dark-navy rounded chip next to a live-typed 'Kolbo.AI' wordmark — both mirror the reference's white-on-dark context |
+| Kolbo K icon, clean cutout | `Youtube\מכונת הזמן\Exports\Thumbnail\assets\k-icon-clean.png` | ⚠️ white-on-transparent — fine for HTML/PIL compositing, but as a *generation reference* on a light background it gives zero contrast and the model redraws the K wrong. Use the black-background source above instead |
+| Kolbo lockup (K + wordmark) | `Youtube\מכונת הזמן\Exports\Thumbnail\assets\kolbo-lockup.png` | ⚠️ this is the **stacked t-shirt lockup** (K on top, small wordmark below) — do not use for a horizontal corner mark, it misreads or gets redrawn wrong when scaled |
+| ByteDance icon | `kolbo-api\assets\Bytedance icon.png` | ❌ **wrong file** — a generic blue bar-chart icon, not ByteDance's real mark. Confirmed by hash, pre-existing bug (not from any recent edit). Filed as `task_f3fab040`. Do not composite this into a "real logo" claim; if a real ByteDance mark is needed, source and verify one first |
+| Zohar — real likeness Visual DNA | id `6a64b8fa5bd226f7e763367b`, name **`zohar`** (single token, so `@zohar` binds) | ✅ correct — use this whenever the ask is "me"/"my face" |
+| Zohar — costumed character DNA | `@zohar_salon` and similar | ⚠️ these are **film-character** DNAs from specific shoots, not his real likeness — never substitute for a "me presenting" thumbnail |
+
+**DNA naming rule that bit us:** the `@Name` prompt tag only binds if it exactly
+matches the DNA's stored `name` field as one token. A DNA named "Zohar (Copy)" can
+never bind via `@zohar` — rename the DNA (`update_visual_dna`) once, don't work around
+it per-prompt.
 
 ## Routing in Kolbo
 
 | Ask | Tool |
 |---|---|
-| One cover | `generate_image` (a single `text_to_image`) |
-| Several options at once | The in-app **Thumbnail Generator** — topic + style + font + aspect, and it fans out 4–8 art-directed variations in one run (it uses Creative Director underneath) |
-| A batch with a locked character or product | `generate_creative_director` with a Visual DNA attached |
-| Cover for a video the user already made | Never crop a frame out of the video — generate a fresh comp. A film frame is exposed for motion, not for a 200px grid |
+| One cover | `generate_image` |
+| Several options at once | The in-app **Thumbnail Generator** — fans out 4–8 art-directed variations |
+| A batch with a locked character or product | `generate_creative_director` with a Visual DNA |
+| Cover for a video the user already made | Never crop a frame out of the video — generate a fresh comp. A film frame is exposed for motion, not for a 168px grid |
+| A second aspect ratio (e.g. 9:16 Shorts after a 16:9 main) | **Generate it natively at that aspect ratio**, do not crop/recompose the other one in HTML. Reuse the same brief, DNA, and reference logo files; write a fresh detailed prompt sized for the new canvas |
+
+**Default workflow, one shot:** `generate_image` (or `_edit`) at **`quality: high`**,
+native target resolution (1024-class for GPT Image 2), passing every real logo file the
+comp needs as `reference_images` plus the matching `@Name`/`#Name` tags in the prompt
+text, and the correct Visual DNA id. Ask for **3 concept variations in the same call
+batch** per the ladder below rather than iterating one image through many small manual
+fixes. Reserve HTML-compositing patches for the rare case a specific mark still comes
+back wrong after 2–3 regenerations with references — it is a fallback, not the default
+path.
+
+**9:16 safe-zone discipline:** keep the hero content (face, gesture, headline, badge,
+logo) inside roughly the **centre 70% of the canvas width**, not edge-to-edge — verified
+by the centre-square-crop test below. State this explicitly in the prompt ("generous
+margin on both sides, nothing touches the left/right edge") rather than fixing it after
+generation.
 
 ## Variation ladder
 
-When producing a set, vary the **concept**, never the words. This is the ladder the in-app
-tool uses, and it is a good default order:
+Vary the **concept**, never the words: bold dynamic · clean minimal · vibrant saturated ·
+cinematic wide · close-up dramatic · typography-forward · dark moody · flat illustration.
+
+Make **3 thumbnails and 10 titles** per video, and design them *before* shooting. Top
+creators spend ~30% of their effort on packaging versus ~5% for small channels. **[B]**
+
+---
+
+## What now reads as dated or "AI slop"
 
-1. Bold dynamic — high contrast, dramatic light, scroll-stopping energy
-2. Clean minimal — one focal point, premium negative space
-3. Vibrant saturated — rich colour, maximum visual impact
-4. Cinematic wide — epic scale, movie-poster feeling
-5. Close-up dramatic — intense subject detail, emotional impact
-6. Typography-forward — the text is the hero, graphic art direction
-7. Dark moody — deep shadows, selective highlights
-8. Flat bright illustration — playful shapes, bold outlines
+**Correct the premise first: YouTube has NOT demoted AI thumbnails.** Its synthetic-content
+policy explicitly exempts "using generative AI tools to create or improve a video outline,
+script, **thumbnail**, title, or infographic." **[A]** Anyone claiming the algorithm
+penalises AI thumbnails is repeating a fabrication. The damage is **competitive and
+trust-based**, and it is real:
 
-## Faces
+- **The models converge — you disappear into sameness.** 700 generation trajectories
+  through image-model feedback loops all collapsed to just **12 dominant motifs**; the
+  authors call the result "visual elevator music." **[A]**
+- **The gloss is the tell** — see the saturation note above. **[A]**
+- **AI faces homogenise.** Diffusion models render same-demographic individuals as
+  near-identical across professions. Your AI face looks like everyone's AI face — which
+  is exactly why a real Visual DNA built from real photos matters. **[A]**
+
+**Avoid — evidence-backed:** exaggerated open-mouth shock · over-saturated HDR/clarity ·
+plastic AI skin · piles of arrows and circles (one or two maximum) · **imagery not
+actually in the video** (policy risk *and* it loses the watch-time test) · AI likeness of
+real public figures (YouTube likeness detection is live for all YPP creators).
+
+**Avoid — credibility damage in tech/AI niches:** glowing brains, robot-hand-touching-
+human-hand, circuit boards, binary code, humanoid robots, the default blue "tech" wash.
+Each is sci-fi-derived and misleading about what AI actually is. For a channel whose
+credibility *is* the product, these are self-harm.
+
+**Avoid on taste, no data — don't cite numbers for these:** golden particle dust and
+bokeh, radial speed lines, neon collage, holographic UI overlays, floating tech icons,
+gradient-mesh purple/blue backgrounds, hexagon grids, generic fantasy creatures.
+
+---
+
+## Tech / AI-tool / filmmaking channels
+
+The grammar here is the **opposite** of the MrBeast formula, and that is the point.
+Restraint reads as authority: controlled product photography, matte dark grounds, one
+accent pulled from the subject itself, minimal or no text, a composed direct-to-camera
+expression rather than a reaction. **The absence of arrows and explosions is the brand
+signal.** For a channel selling expertise, the thumbnail is a claim about whether you
+know what you are talking about.
+
+What works specifically:
+- **The tool logo as the recognisable object** — high search intent, heavily used;
+  differentiate through treatment (scale, lighting, physical staging), not by dropping it.
+- **Before/after** — genuinely strong for AI filmmaking because the transformation *is*
+  the promise and it is honest. Skip the arrow; let the two images do the work.
+- **The interface as hero** — a real UI cropped tight to one striking element signals
+  "actual tutorial, not hype." Under-used.
+- **Face + artefact** — face at ~⅓ frame, closed-mouth focused, beside the generated
+  frame. Satisfies both the face finding and "show the thing."
+
+---
+
+## Hebrew and RTL
+
+> **No evidence base exists for any of this.** There is no measured data on Hebrew vs
+> Latin thumbnail text, Hebrew legibility at 168px, or RTL thumbnail composition. What
+> follows is typographic reasoning plus Israeli foundry commentary. For a Hebrew channel
+> this is worth settling empirically — ship a text-right/subject-left variant against a
+> mirrored one and let YouTube's own test decide.
+
+- **Hebrew has no ascenders or descenders and a uniform x-height.** Latin words have a
+  ragged silhouette that aids word-shape recognition at small size; Hebrew words are
+  near-rectangular blocks. **Hebrew needs more size, more weight and more letterspacing
+  than Latin to hit the same legibility at 168px.** Budget for it. **[C]**
+- **Fonts:** **Ploni** (AlefAlefAlef) is purpose-built bilingual — Hebrew plus Latin
+  designed to sit together without either overshadowing the other, which is exactly the
+  problem when a tool name like `Seedance 2.5` sits inside a Hebrew phrase. **Heebo** and
+  **Rubik** are strong free alternatives with real bold weights.
+- **Mirror the layout, not the logos.** RTL readers scan a mirrored F-pattern entering
+  from the **top-right**, so the natural composition inverts: **text block right, subject
+  left**, gaze pointing right-to-left toward the text. Latin brand marks and numerals stay
+  LTR regardless. **[B]**
+- **Language split:** Hebrew for the emotional/promise word, Latin for the tool name — the
+  tool name is the search-intent anchor and the audience already reads it in Latin.
+- **Per-language thumbnails exist** but are gated behind multi-language audio tracks: you
+  need at least one added audio track before a localised thumbnail can attach to it. **[A]**
+
+---
 
-If the channel has a host, lock them with a **character Visual DNA**
-(`workflows/visual-dna.md`) so every thumbnail in the series is the same person. Expression
-is the payload: shock, delight, disbelief, triumph. A neutral face is a wasted thumbnail.
+## Honest limits
 
-## Phone-shot thumbnails
+Thumbnail design is a craft with weak empirical foundations. A study of 2,400 news
+thumbnails across 21 visual features found **few statistically significant correlations
+with engagement at all** **[A]**. The breakout data above describes what winning
+thumbnails *look like*, not what *caused* the win. Genuinely unsettled: rule-of-thirds vs
+centred (no data either way), and saturation levels (creator orthodoxy vs the aesthetics
+literature). Do not present either as settled.
 
-A "raw / authentic" cover is a real style — a phone-shot frame with big type over it beats
-a polished render for vlog and UGC channels. Build the plate from
-`workflows/ugc-smartphone.md`, then apply the text rules above unchanged. The type stays
-graphic and deliberate even when the photo is deliberately casual.
+A/B testing needs ~10k impressions per variant to mean anything. Below that, just replace
+the thumbnail outright.
 
 ---
 
 ## Checklist
 
-- [ ] One subject, 40–70% of frame, tight crop
-- [ ] Subject separates hard from the background
-- [ ] ≤ 4 words, quoted verbatim, everything else forbidden
-- [ ] Text has an outline or backing bar, sits off the face
-- [ ] Correct aspect, critical content inside the safe area
-- [ ] Checked at 200px (and centre-cropped first, for 9:16)
-- [ ] No watermark, no stray words, no gibberish letters
+- [ ] One idea, nameable in under a second at 168×94
+- [ ] Face ~⅓ of frame, **one legible emotion, closed-mouth**, eye contact
+- [ ] Greyscale test: subject still separates
+- [ ] ≤3 colours, restrained saturation, checked on light *and* dark themes
+- [ ] 0–3 words, outlined, ~12–15% frame height, off the face, never bottom-right
+- [ ] Nothing in the bottom-right quadrant or bottom 40px
+- [ ] Real logos composited from files, never model-rendered
+- [ ] Nothing in the image that isn't in the video
+- [ ] Exported 3840×2160; verified at 168×94 (and centre-cropped first, for 9:16)