koubou（koubou 搜索）

Koubou: App Store Screenshot Generator

Generate professional App Store screenshots using HTML/CSS templates with 100+ real device frames, xcstrings localization, and pixel-perfect Apple dimensions.

Screenshots are advertisements, not documentation. Each slide sells one feeling, outcome, or pain-point solution.

User instructions always take priority over defaults in this skill.

Creative direction is open by default. Use the rules in this skill to enforce impact, readability, and canvas-aware scale, not to make every set look the same.

Rule Hierarchy

Treat the guidance in this skill in this order:

1. Hard constraints — readability, scale, thumbnail clarity, truthful product marketing, and real canvas awareness
2. Strong defaults — variety, asymmetry, overlap, rhythm, and avoiding generic layouts
3. Taste preferences — anti-slop heuristics and stylistic defaults that can be broken when a stronger idea clearly improves the work

If breaking a default produces a better screenshot set, break it deliberately and explain the tradeoff to yourself before generating.

Reference files (read as needed, not upfront):

• setup.md — Installation and HTML runtime setup
• design-guide.md — Design principles, copywriting rules, CSS rules, HTML template examples
• style-intake.md — How to inspect the app's real visual language before asking questions
• style-interview.md — Short, high-signal style questions and conversation patterns
• yaml-reference.md — YAML config format, localization, assets, devices, sizes
• capabilities-reference.md — Full koubou capabilities (content mode, highlights, zoom, gradients)

Phase 1: Setup (silent, automatic)

Check if kou is available first. Do not reinstall or touch Python packaging if kou already works.

Preferred order:

kou --version 2>/dev/null

If kou --version succeeds:

• treat koubou as installed
• do not run pip, pip3, or playwright install
• do not try to "upgrade" or "fix" the user's environment proactively
• always run kou setup-html before generating, because this skill is HTML-only and the command is designed to be safe to repeat

If kou --version fails, read setup.md and follow the least invasive path for the environment.

For HTML rendering support:

• always use kou setup-html
• do not skip it just because HTML worked in a previous run
• do not run playwright install chromium directly

Only inform the user if setup fails. Never mutate a working installation.

Phase 2: Read The App, Then Ask

Do not begin with a generic style questionnaire. Inspect the app's real visual language first, then ask only what is still missing or ambiguous.

Read style-intake.md before asking style questions. Follow this order.

2.1 Style discovery (mandatory before style questions)

Inspect local project sources first. Search automatically before asking the user for visual direction.

Prioritize these sources:

1. Assets.xcassets, app icons, illustrations, logos, color assets
2. Existing screenshots and marketing folders:

- .maestro/ - screenshots/ - AppStore/ - fastlane/screenshots/ - marketing/

1. Product docs and positioning:

- README* - CLAUDE.md - launch plans, landing-page copy, docs folders

1. Design tokens and UI code:

- CSS variables - SwiftUI colors, gradients, materials, typography choices - web theme files and shared design constants

1. Prior App Store artifacts if they exist:

- previous templates - old screenshot campaigns - localized marketing assets

Extract and summarize at least these signals:

• dominant colors and accent colors
• overall contrast: dark, light, mixed, muted, vivid
• UI density: airy, balanced, dense
• shape language: sharp, soft, rounded, card-heavy, flat
• iconography and illustration style
• typography direction if visible
• copy tone: calm, technical, playful, premium, warm, urgent
• whether the app feels calm, technical, warm, energetic, polished, playful, etc.

If the app style is clear:

• summarize the detected signals to the user in plain language
• confirm that you will build from that direction instead of restarting from a generic template language

If the signals are weak, missing, or contradictory:

• say what you found
• ask focused follow-up questions about the uncertainty instead of asking a broad design questionnaire

2.2 Intent interview (short, high-signal)

Read style-interview.md and ask only the minimum needed. Keep it conversational. Do not dump every question at once.

Required style questions when still needed:

1. What feeling should the campaign project: premium, playful, editorial, utilitarian, technical, warm, etc.?
2. Are there App Store references, brands, or screenshot sets that should influence the direction?
3. What is visually forbidden: loud gradients, glassmorphism, dark-only, card grids, overly playful styling, etc.?
4. What visual trait from the app itself must still be recognizable in the screenshots?

Optional questions only when still unresolved:

1. Preferred type direction or font personality
2. Light, dark, or mixed bias
3. Density preference: more air vs more information
4. Whether the set should feel tightly consistent or visibly varied across slides

2.3 Product information (still required)

After style discovery starts, collect the remaining production inputs naturally:

1. App: Name, what it does (1 sentence), main value proposition
2. Screenshots: Where are the app captures? Search automatically first:

- If found, show what you found and confirm - If not found, ask how the user generates them - Prefer clean simulator captures that show the app UI clearly. If the user can choose, prefer recent 6.1-inch iPhone captures as the source material

1. Features to highlight: Prioritized list of features/benefits (recommend 3-5). Each slide = 1 feature
2. Hero assets: App icon or other brand asset. Search automatically before asking
3. Slide count: How many screenshots (recommend 3-5 to start, Apple allows up to 10)

Optional only if relevant:

1. Device: Default is iPhone 16 Pro - Black Titanium - Portrait. Only ask if iPad/Mac/other makes sense
2. Localization: Only if the project already has xcstrings or multiple language support
3. Extra assets: Floating UI elements, badges, supporting illustrations
4. Additional constraints: Required claims, forbidden styles, or marketing constraints

2.4 Style decision (mandatory before HTML)

Before drafting templates, explicitly decide and internally lock these six items:

• brand signals detected
• chosen campaign style
• copy voice
• background system
• device composition rhythm
• variation plan across slides

Each item must be justified by either:

• app evidence from style discovery, or
• direct user feedback from the interview

If you cannot justify one of these, keep asking focused questions before designing.

Derived (do NOT ask unless blocked)

• Background gradients or textures from brand colors and UI mood
• Layout distribution (hero → feature-top → feature-bottom → alternating)
• Headline copy from features and value proposition
• Secondary palette from the app's actual palette
• Canvas class from project.device + project.output_size via kou inspect-frame
• Whether the app name should appear at all. Default: omit it unless it adds real brand value at readable size
• Whether the app icon should appear. Default: use it sparingly on hero or closing slides, not as a tiny decorative marker

Principle: if context is available locally, use it before asking. Style questions come after inspection, not before.

Phase 3: Generate

Read design-guide.md before generating templates. Read yaml-reference.md before writing config.

1. Create working directory (e.g., AppStore/ or wherever makes sense for the project)
2. Run kou setup-html
3. Read project.device and project.output_size from the YAML
4. Run kou inspect-frame "<device>" --output-size <size> --output json and use that geometry before writing CSS
5. Draft the narrative arc and 2-3 headline/subtitle options per slide before touching layout. Pick the strongest one first. Copy quality comes before CSS
6. Translate the style decision into a campaign brief:

- what the set should feel like - what motifs are allowed - what motifs are banned - how much variation the first 5 slides should show

1. Create templates/ with HTML templates — minimum 3 distinct layouts, and the first 3 slides must not repeat the same composition archetype
2. In every HTML template, add data-kou-id and data-kou-role to these elements before writing any CSS: Without these, the layout sidecar elements array will be empty and post-render QA is blind. See design-guide.md template examples for correct annotation patterns.

- Main headline: data-kou-id="headline" data-kou-role="headline" - Supporting copy: data-kou-id="subtitle" data-kou-role="supporting" - Primary device image: data-kou-id="device" data-kou-role="device" - Feature titles/subs (closing/feature slides): data-kou-id="feature-N-title" data-kou-role="feature-title", data-kou-id="feature-N-sub" data-kou-role="feature-sub"

1. Create config.yaml with koubou config (read yaml-reference.md for format)
2. Use CSS that adapts to the canvas and copy length: use vw or clamp() deliberately, prefer CSS Grid, overlap, and absolute-positioned layers, and verify the final computed text scale on the real canvas
3. Before writing HTML: plan text/device zones for each slide (which area owns what percentage of the canvas). Do not start CSS until zones are clear
4. Run: kou generate config.yaml --output json
5. For each generated HTML screenshot, read layout_path from the JSON result and inspect the sidecar before deciding the layout is acceptable
6. Post-render QA (mandatory — do not skip):

- Review each generated slide visually - Open each *.layout.json sidecar referenced by layout_path - Check against the rejection checklist in design-guide.md - Check that the rendered set still reflects the app style you discovered instead of generic App Store defaults - Use layout JSON only for objective geometry: positions, occupied space, proportions, and mathematical overlaps - If elements is empty for an HTML slide that should be measured, treat that as missing or broken annotation and fix the template - Verify: no emoji icons, no identical card grids, no decentered devices, no text below minimum scale, no unintentional device cropping (device top/notch must be visible on hero slides; screen content must be readable) - Apply the logo-swap test: would these slides work for a competitor? - If any slide fails, fix and regenerate before showing to the user

1. Open output folder: open <output_dir>
2. Ask if the user wants adjustments — iterate on specific slides without regenerating everything

Iteration Rules

• When user asks to change a specific slide, only modify that template + config entry
• When user asks for a global style change (colors, fonts, mood, density, boldness), update all templates
• Re-run kou generate config.yaml --output json after changes so the new layout_path values and sidecars stay in sync
• Use kou live config.yaml if user wants real-time preview while editing
• If a slide feels small, first increase scale or switch layout. Do not hide the problem with extra gradients or labels
• If copy forces tiny text, rewrite the copy or choose a more suitable layout; never accept a timid slide because "the text had to fit"
• Keep variety high: the goal is campaign consistency, not layout repetition
• Do not stop after the first successful render if the output still violates the design rules or rejection checklist
• If the user asks for a mood change, reinterpret the whole set inside the quality limits instead of defending the previous defaults
• If screenshot and layout JSON disagree, trust the rendered screenshot for taste and investigate the annotation or render timing instead of forcing a layout decision from stale geometry

Hard Rules

Hard constraints

• Never use emoji as icons — use CSS shapes (accent bars, dots) or text-only
• Never use banned copy phrases — "revolutionary", "seamless", "unlock", "game-changing", etc.
• Apply the logo-swap test — if a competitor's name would fit, the set is too generic
• Every HTML template must have data-kou-id annotations — at minimum: headline, subtitle, and device. If elements in the layout sidecar is empty after generation, annotations are missing — fix the template before QA passes. Do not present unannotated slides to the user.

Scale

• On tall iPhone portrait (iPhone6_9, iPhone6_7), hero headlines must be >=10vw, side-layout headlines >=10vw, subtitles >=4.5vw
• Use vw or clamp() only if the computed size still clears the minimum on the target canvas
• On contrast or closing slides, the content cluster must occupy 60%+ of canvas height

Strong defaults

• Do not use the same upright, centered phone composition on consecutive slides
• Do not leave large empty bands (>15% canvas height) between headline and device
• Avoid combining rotate() + translate() on centered devices unless the visual center still feels intentional after render
• Do not use a simple centered flex column when the slide needs overlap, edge-breaking crops, or layered composition
• Plan text/device zones before writing CSS (see design-guide.md Device-Text Zone Planning)
• If you use card grids, stagger, emphasize, or simplify them unless a strict grid genuinely serves the concept

Content

• Do not add tiny top labels, category labels, or app-name headers by default
• Do not render the app name as small decorative text unless clearly readable and strategic
• Do not use app icons as tiny corner decorations; if used, they should be a real brand element
• Do not use awkward, literal, or unnatural headlines just to be short
• Do not start layout work before the slide narrative and copy are coherent

Process

• If the first 3 slides do not already feel App Store-ready, keep iterating before presenting them
• Do not present the first technically successful render as final — always review against the rejection checklist in design-guide.md
• Do not let the rules flatten the creative direction. A strange but strong composition is valid if it stays readable and high-impact
• Defaults can be broken when the final result is stronger, clearer, and more specific to the app
• Do not start from the generic fallback of dark gradient + centered phone + big white headline unless the app genuinely supports that language or the user explicitly wants it
• You should be able to explain to yourself why the campaign looks like this app and not a competitor with the same feature list

Key Technical Details

How assets work in HTML templates

In template HTML, reference assets with {{asset_name}}. In the YAML config, map asset_name to a file path under assets:.

Koubou automatically pre-renders each image asset with the configured device frame before passing it to the HTML template. The template receives a composited image (screenshot inside device frame) — it just places it with <img src="{{asset_name}}">.

To disable frame for a specific screenshot: set frame: false in its YAML definition.

Template variable substitution

• variables: in YAML → {{key}} in HTML → localizable text (extracted to xcstrings)
• assets: in YAML → {{key}} in HTML → image file paths (pre-rendered with device frame)

Layout JSON for HTML screenshots

For HTML screenshots, Koubou can emit a compact sidecar JSON with measured layout geometry.

• Use data-kou-id on any element that should be measurable
• Use data-kou-role only when the role helps the model interpret the element
• Keep annotations minimal and structural, not exhaustive
• Good defaults: annotate the main headline, supporting copy, and primary device or hero image

Example:

<h1 data-kou-id="headline" data-kou-role="headline">{{headline}}</h1>
<p data-kou-id="subtitle" data-kou-role="supporting">{{subtitle}}</p>
<img data-kou-id="device" data-kou-role="device" src="{{app_screenshot}}" alt="">

Generation workflow:

• Run kou generate config.yaml --output json
• Read layout_path from the command output
• Open the referenced *.layout.json

Interpretation rules:

• Geometry fields are normalized ratios from 0..1
• elements contains only annotated nodes
• overlaps contains only mathematical box intersections
• Use this file to understand layout facts, not to decide taste
• Do not invent subjective rules such as "too small" from the JSON alone; combine the geometry with the actual screenshot review

Output structure

{output_dir}/{language}/{device_name}/{screenshot_id}.png

Non-localized projects skip the language directory.

Device frame names

Use exact names from kou list-frames. Common ones:

• iPhone 16 Pro - Black Titanium - Portrait
• iPhone 16 Pro Max - Black Titanium - Portrait
• iPad Pro 13 - M4 - Space Gray - Portrait

Search with: kou list-frames "iPhone 16"

Frame inspection for layout decisions

Use kou inspect-frame "<device>" --output-size <size> --output json to get:

• real frame size
• screen bounds and screen bbox
• safe margins
• coverage ratio
• orientation
• canvas_class

Use this data to choose typography scale, crop aggressiveness, and layout density.

Output sizes (App Store dimensions)

Custom: output_size: [1320, 2868]