AI plan · v1 · October 2026
How a two-person character studio builds with Claude.
One character bible feeds every pipeline: multilingual commerce, print animation, a puzzle game and customer service. Heeseung’s hand-carved art stays human-made; Claude writes, plans, builds the tools and checks the work, and a person approves everything that ships.
01
One character bible, every pipeline
Everything starts from a single, versioned document that describes our world: who each character is, how they move and speak in each language, what they would never do, and the visual rules of a relief print (limited palette, mirrored carving, visible grain).
What it holds
- Character sheets: personality, habits, relationships
- Voice guide per language (KO / JA / EN / zh-TW) with do / don’t examples
- Glossary: product names, character names, fixed translations
- Visual rules for print layers, colors and framing
How Claude uses it
- Loaded once as a cached system prompt, so every call shares the same voice at low cost
- Kept as Markdown in our repository; changes go through review like code
- A 30-example evaluation set is re-run whenever the bible or prompts change, to catch drift in voice
Why: the hardest part of growing a character brand across languages is staying recognisably us. Putting the voice in one place makes every new channel cheaper and more consistent.
02
Multilingual commerce
Oct–Dec 2026 · first deadline: Taipei SIF Illustration Fair, Oct 23–26
Export the catalog Claude Code
A small script pulls our 50+ listings into one structured file (name, materials, size, price, photos).
Translate in brand voice Claude API
Listings and booth cards go to EN / JA / zh-TW with the bible as a cached prompt; bulk runs use Message Batches.
Draft from photos Claude vision
For new items, Claude reads product photos plus our spec notes and returns title, description, size table and alt text as structured output.
Approve CEO
Heeseung reviews a side-by-side sheet and approves or edits each text. Nothing is published without it.
Publish & print Claude Code
Approved text flows into the store, the website and print-ready booth cards.
03
Print animation & animated emoticons
Short, wordless films (15–30 seconds) in which our characters move the way a print is made: layer by layer, with the grain of the paper showing. Wordless means one film works in every country; only captions and titles change. The same pipeline produces an animated KakaoTalk emoticon set from our existing 42 expressions.
Ideas Claude API CEO picks
Each week Claude proposes 10 loglines from the bible and the seasonal calendar (fairs, holidays, new products). Heeseung picks two.
Beat sheet & shot list Claude API
Claude turns a logline into beats and a shot list in structured form:
shot, seconds, framing, action, layers, sound.Key art Heeseung
Heeseung carves and prints the key frames and separate layers (character, props, background). The art is always hand-made.
Motion assembly Claude Code
Our own tool, built with Claude Code, rigs the scanned print layers as cut-outs and animates them on the shot list’s timing (limited animation, 12 fps), then renders with ffmpeg. We do not use generative video for our characters.
Frame check Claude vision
Claude compares sampled frames with the shot list and bible: missing layers, wrong colours, continuity, caption safe areas. Issues come back as a checklist.
Captions & release Claude API Approve
Titles, captions and descriptions in KO / JA / EN / zh-TW; we approve and publish to Shorts, Reels and our site.
Shot list format
// one shot from a 20-second film, produced by Claude, read by our assembly tool
{ "shot": 3, "seconds": 2.5, "framing": "medium, eye level",
"action": "the dog lifts the book with its nose; pages flutter",
"layers": ["bg_bookshelf", "dog_body", "dog_head", "book_open"],
"motion": { "dog_head": "tilt +8deg ease-out", "book_open": "rise 40px" },
"sound": "paper rustle" }
Targets
- Pilot: 4 films by end of Q2 2027
- Then 2 films a week
- Animated emoticon set prepared from the existing 42 expressions
Measured by
- Hours of human work per film
- Frame-check issues caught before release
- Views and follows per language
Why this way: our value is the hand-made print. Claude takes over the writing, planning, tooling and checking around the art, so a two-person team can publish animation weekly without outsourcing and without imitating our own style with a generator.
04
Game: “Print & Stroll” (working title)
A cozy puzzle game for phones and the web, built on how relief printing actually works: what you carve comes out mirrored on paper. Players stroll through a small neighbourhood with our characters, find scenes with something missing, and carve a block so that the print fills the gap.
Core loop (3–5 minute sessions)
Mechanics
- Rotate, flip and place carved pieces on a grid; the print is the mirror image
- Ink colours stack in layers, like a real multi-block print
- Difficulty grows with grid size, layers and limited moves
Connection to the business
- Completed sketchbook pages unlock real sticker designs
- QR codes on our goods unlock special stamps in the game
- A tablet demo becomes a booth attraction at our fairs
Where Claude comes in
| Area | Claude’s role | Human role |
|---|---|---|
| Building | Claude Code writes the game (TypeScript, Phaser, installable web app; mobile wrapper later), the level editor and automated tests | Jiho reviews and plays every build |
| Levels | Claude proposes puzzles as JSON in bulk (Message Batches); our solver keeps only levels with exactly one solution at the target difficulty | We curate the best levels and order them |
| Story & lines | Scene stories and character lines written from the bible, then localized to JA / EN / zh-TW | Heeseung approves every line; all text is fixed before release, nothing is generated live |
| Art pipeline | Claude Code builds the scan → clean-up → sprite-sheet tools | All art is Heeseung’s prints |
| Playtests | Claude groups tester feedback and play data into themes and suggested fixes | We decide what to change |
Level format
// a level proposed by Claude; kept only if the solver finds exactly one solution
{ "id": "walk-012", "grid": [6, 6], "layers": 2,
"target": "bird_nest_on_branch", "pieces": ["L3", "T4", "dot", "dot", "bar2"],
"rules": { "mirror": true, "max_moves": 9 },
"difficulty": 3, "unlock": "sticker_nest" }
Milestones
- Q1 2027 — web prototype with 10 levels
- Q2 2027 — tablet booth demo, 50 levels
- H2 2027 — release on web and mobile, 120 levels in 4 languages
Measured by
- Day-1 and day-7 return rate
- Share of puzzles solved without hints
- Store visits and QR unlocks coming from the game
05
Customer & wholesale inquiries
Claude reads an incoming question in any language, looks up stock, shipping rules and stockist information through tools, and drafts a reply in the customer’s language. A person reads, edits and sends every reply; nothing is answered automatically. We collect only what a reply needs.
06
Roadmap & metrics
| When | What ships | Claude features |
|---|---|---|
| Live now | Website built with Claude Code; EN / KO / JA versions translated by Claude | Claude Code |
| Oct–Dec 2026 | Character bible v1; catalog and booth cards in EN / JA / zh-TW | API, prompt caching, Batches |
| Q1 2027 | Listing drafts from photos; game web prototype (10 levels) | Vision, structured outputs, Claude Code |
| Q1–Q2 2027 | Inquiry assistant with human approval | Tool use |
| Q2 2027 | First 4 print-animation films; animated emoticon set; game booth demo | API, vision, Claude Code |
| H2 2027 | Weekly films; game release on web and mobile | All of the above |
Expected usage at full run is roughly ten million tokens a month, most of it level generation and frame checks. Program credits would let us build and test these pipelines on real volume before overseas sales and the game pay for them.
07
Guardrails
Human approval
Every text, film and level that reaches a customer is approved by a person. The game ships with fixed, reviewed text only.
Hand-made art
Character art is always Heeseung’s prints. We do not use image or video generators to imitate it.
Minimal data
The inquiry assistant sees only what a reply needs; no customer data is used for anything else.
Versioned prompts
The bible and prompts live in our repository with an evaluation set, so changes are reviewed and tested like code.