01 Overview
A game's worth of 3D assets, and an artist who should only be making the calls.
Star Online is our own Unity game. Asset Workshop is the pipeline that makes its props and buildings: the agents do the measuring, the repetitive work and the bookkeeping, and the artist keeps control of the creative decisions.
01.1 The problem
The steps from concept art to Unity were never the hard part. Carrying the asset between them was.
The work was always the same steps. Doing them by hand meant moving files from one tool to the next and checking every model by eye.
| What was going wrong | What we did about it | |
|---|---|---|
| Automation problem Every asset went from concept art to model to Unity by hand: files carried between tools, and each model checked by eye, which is very hard to do well. ~3 days per asset From idea to a file Unity could use, with AI already in the loop. | Agents run the steps and carry the handoff between them; the artist makes the calls at evaluate and polish. | See the work |
01.2 The timeline
Started in August 2026; four of the game's eight zones done by November.
Built for our own game, and still running.
- 2026August
Work on Asset Workshop starts.
- October
Moves out of the Unity project into its own repository.
- November
280 assets exported to Unity; four zones done, and still running.
01.3 Context
Our own game, one person on it, and no way of working before this one.
What the pipeline started from.
Our own game
Star Online is our own project, so everything it makes can be shown.
One person
Chris built the pipeline and is the artist making its calls. He's also the game's level designer, with an upstream workflow where he sets the art direction.
No before
The pipeline started with the game, and it used AI from the start. Before this workflow, the same steps were done by hand.
The aim
"I want a pipeline that can be trusted and is reliable."
02 Automation solution
Every step written down, and every creative call left to the artist.
Each step is a written skill the agent follows, backed by shared tools, git commits and persistent memory of each asset family's rules. Nothing generated is ever overwritten, so the work is repeatable and every asset can be traced.
02.1 The pipeline
The whole pipeline, from a zone's references to the file Unity imports.
The models track from above: which tool runs each step, and the two places the artist decides. The sections after it are the parts, the steps and the calls behind them.
- Per zone
- InspirationReferences from the artist
- Art direction kitCodex: layout, palette, overviews
- Per asset
- Concept blueprintCodex: front, back and a brief
- 3D attemptsMeshy, untextured
- EvaluateClaude Code scores each attempt
- The artist decidesTexture or regenerateNot closeRegenerateAnother attempt in Meshy
- Texture
- TexturedMeshy
- Finish in Blender
- FittingClaude Code: scale and facing
- Polish planClaude Code, before and after
- The artist approvesBefore the plan is applied
- Approve
- ExportFBX, texture, JSON to Final/
- UnityThe importer reads Final/
02.2 How it fits together
One art direction kit, read by every step after it.
Inspiration and references become an art direction kit for each zone of the game: a layout plan in metres, a colour palette and bird's-eye concept overviews. Every later step reads it, so the assets stay consistent in style from the first blueprint to the final export.
02.3 The parts
Nine parts, from a zone's art direction to the file Unity imports.
What the pipeline is made of, in the order an asset meets it, and who runs each part.
| The part | What it does | Where it runs |
|---|---|---|
| The art direction kit | A layout plan in metres, a colour palette and bird's-eye concept overviews, one kit per zone. | Codex, from the inspiration and references. |
| The concept blueprint | Matched front and back views, plus a brief with the asset's real-world size, triangle budget and import settings. | Codex's image generation, per prop or building. |
| The 3D attempts | Untextured models from the blueprint views, sent into the asset family's Blender file. | Meshy, uploaded by the artist. |
| The evaluation | Measures each attempt against the blueprint, the brief and the in-game camera, scores it, and recommends texturing or regenerating it. | Claude Code in Blender; the artist decides. |
| The fitting | Files each textured model, scales it to the brief and checks which way it faces. | Claude Code in Blender. |
| The polish plan | Sharper edges, hidden faces removed, texture sharpening, colour grading to the zone's palette, texture size: one plan, shown with before and after images. | Claude Code, once the artist approves. |
| The export | Technical checks, then an FBX, a resized texture and a JSON record in a Final/ folder, which Unity's importer reads directly. | Claude Code, into the Unity project. |
| The memory | Each asset family's rules, kept between sessions alongside git commits. | Shared by every step. |
| The other tracks | Foliage, trees, terrain textures and sound effects, each its own track of skills after the zone's art direction, each skill naming the next. | Image steps in Codex, Blender steps in Claude Code. |
02.4 The automation process
Six steps from idea to Unity, and the artist makes the call at two of them.
In the evaluate step, Claude scores each attempt and recommends texturing or regenerating it; the artist decides. It catches what an eye misses: folded hidden triangles that bake into black smudges under the game's lighting. In the finish step, it proposes one polish plan and applies it once the artist approves, with before and after images.
- Art direction kit
- Concept blueprint
- 3D generation
- EvaluateBack to 03
- Finish and polish
- Export to Unity
Caught at 04 Evaluate
The character's height model sits in the Blender file during evaluate, so every attempt is checked against it.
Generated models came out at the wrong scale next to the rest of the game.
- The result
- Scale is checked visually on every attempt; the problems were minor.
- Who, and where
- Chris and Claude, in Blender
02.5 Automation decision
Variations are drawn cheap before anything is modelled, and one step stays by hand on purpose.
Two calls shaped where the agents stop. Both trade a little manual work for a cheaper, surer result.
| The call | The easy answer | What we chose | What it cost, and what it bought |
|---|---|---|---|
| Where the variations get made | One tool for everything: generate every variation in Meshy. | Codex draws variations until one is close to the vision; Meshy only models the chosen blueprint. | An extra image step before any model exists. Exploration at Codex's cost; doing it in Meshy would be very expensive. |
| Who sends the views to Meshy | Have the agent call Meshy's API. | The artist uploads the views and sends the untextured tries into Blender by hand. | "The only thing that's stopping us right now from making this fully agentic." The API returns a model already textured; by hand, each untextured try is tested and evaluated and retried before any texturing. |
03 Field notes
What a few months of running it taught us.
What came of the work that the steps don't show.
Renaming
Variations get renamed later, and a rename duplicated the import in Unity. Each model now gets an ID, and its record carries the old ones so Unity renames in place.
The artist's calls
Overrules happen mostly on art direction: "There are calls that only I can make."
For another studio
"Do it right away, don't wait."
04 Team and stack
One person across automation.
Built on Unity. Made with Claude Code, Codex, Meshy, Blender, and Git.
- Built on
- Unity
- Made with
- Claude Code · Codex · Meshy · Blender · Git
Chris Javier Oliveros
AI Automation Engineer
Built Asset Workshop alone: the written skills, the shared tools and the memory each step runs on. Makes the art direction upstream as the game's level designer, and every creative call in the pipeline as its artist.
- One asset goes from idea to Unity in 30 minutes at most, down from about three days when the same steps were done by hand.
- 280 assets exported to Unity by 2026-11-14 (171 models, 73 foliage, 26 trees, 10 zone files), with four of the game's eight zones done.
05 Impact
About three days an asset, by hand. Now 30 minutes at most, and 280 assets in the game.
Each row is how the work ran before and after. Nothing was timed with a tool: the time per asset is Chris's estimate; the asset count is read from the pipeline's exports.
Time per asset
idea to a file Unity imports30 minutes at most
Before
About 3 daysAssets in Unity
August to November 2026; 171 models, 73 foliage, 26 trees, 10 zone files280
Before
NoneZones done
August to November 20264 of 8
Before
None