t1k:minimax-design
| Field | Value |
|---|---|
| Module | t1k-multimodal |
| Version | 2.10.46 |
| Effort | medium |
| Tools | Bash, Glob, Grep, Read, Write |
Keywords: game art generation, game asset generation, generate image with minimax, icon set generation, image generation mcp, mcp minimax, minimax design, minimax-design, transparent background image, ui asset generation
How to invoke
Section titled “How to invoke”/t1k:minimax-design<what to generate> [project] — generation, batch, reskin, troubleshootMiniMax Design MCP — image generation for game art and UI
Section titled “MiniMax Design MCP — image generation for game art and UI”minimax-design is an MCP server that drives the MiniMax Design desktop app: it generates images
into a project the user has open, bills credits to the account signed in to that app, and saves
every result as a version inside the project. This skill is the tool procedure. Project conventions
(which folder an art style lives in, a particular game’s naming) belong in project memory, not here.
Executed by whichever session actually holds the mcp__minimax-design__* tools. Whether that
includes a sub-agent depends on how the server is registered: when it is registered at user
scope, Claude Code lists the namespace in each spawned sub-agent’s own tool set, and a sub-agent
can run a generation lane (measured on Claude Code 2.1.316: six general-purpose agents at
spawnDepth: 1, 222 MiniMax tool calls, every agent finished: true; the is_error results in
that sample were ToolSearch probes, not generation calls). When the connection belongs to one
session only, the tools are absent from the child and the lane has to run inline. Do not assume
either direction: probe with one read-only call (list_projects) in the child before handing it a
batch, and fall back to inline if the namespace is missing. Contact-sheet review (step 5) stays
delegated through the router vision role either way.
The one rule: credits are real money
Section titled “The one rule: credits are real money”Every execute call bills. Never resubmit a task — poll it with task_wait on the SAME
taskId. Never call confirm_cost on your own behalf: the cost prompt belongs to the user in
the MiniMax app, and confirm_cost may be called only after the user explicitly agreed to that
amount in the conversation. State a credit estimate and smoke-test before any batch.
Decision tree
Section titled “Decision tree”| Intent | Read |
|---|---|
| “Generate an icon / asset / UI screen” | Steps 1–6 below |
| Batch, level sets, reskinned screens, prompt craft, local resize, Figma upload | references/production-playbook.md |
Exact tool arguments, execute/deliver/get_output payload shapes, error codes | references/tool-reference.md |
Something failed (PROJECT_NOT_OPEN, OUTPUT_UNAVAILABLE, INVALID_ARGUMENT) | Gotchas table below |
Setup (once per session)
Section titled “Setup (once per session)”The tools are DEFERRED: load their schemas first, or every call fails with InputValidationError.
ToolSearch select:mcp__minimax-design__list_projects,mcp__minimax-design__open_project,...If they are missing entirely right after the server was registered, the tool list was snapshotted at
session start — reconnect (/mcp) or restart before trying again. The endpoint/port is
machine-specific: check claude mcp list, never assume one.
Workflow
Section titled “Workflow”- Find the project.
list_projectsreturnsprojectId,name,folderPathand anopenflag. Only a project withopen: trueis usable. Never guess between similar names (two near-identically named projects, one open): pick the open one and say which you picked.open_project(projectId)switches; calling it again switches again. - Reuse before generating.
composition_searchbrowses what the project already holds andcomposition_getlists members with theirresourceId/currentVersionId. Generate only what is genuinely missing; imported references are billed on every task that carries them (see Costs). - Read the parameters once.
capability_list(kind="media", operation="generate_image")lists models;capability_get(ids=[...])returns each one’sparamSchema. Default model unless the user names another:generate_image:gpt-image-2.5-sunburst(supports transparent background). Every other model is “use only when explicitly requested”. - Import approved style references.
deliveris synchronous and unbilled; it turns local files intoresourceIds you can pass asreferences. The path MUST be a regular file inside the project’sfolderPath— copy it into a subfolder of the project first. Import once, reuse the ids. - Smoke, then look. Generate 3–4 images covering the hard cases (lowest and highest of a level
set, an odd aspect ratio, one other colour). Then review them the way a human would: build a
contact sheet and run the vision role
(
bun <claudeDir>/scripts/mr-role.cjs call --role vision --file <sheet> --prompt "..."). Show the user and stop for approval before scaling. - Batch in waves, costed. State the estimate before a batch of more than ~10 images or more
than ~2 minutes, and re-check visually between waves — never spend the next wave on unverified
output. Multi-level sets and reskins:
references/production-playbook.md. - Finish locally, then deliver. The model cannot take pixel sizes (see Parameters), so resize
or pad on disk, verify with a final contact sheet, and only then hand off. Results already live
inside the project;
get_output(versionId)returns their absolute local path. Getting them into Figma is a separate hand-off:references/production-playbook.md§ Figma.
Parameters (gpt-image-2.5-sunburst)
Section titled “Parameters (gpt-image-2.5-sunburst)”| Key | Values | Notes |
|---|---|---|
aspect_ratio | 1:1, 16:9, 9:16, 3:4, 4:3, 3:2, 2:3, 5:4, 4:5, 21:9, 2:1, 1:2, 3:1, 1:3 | Required when there are no image references; with references it inherits the nearest ratio of the FIRST reference if omitted |
resolution | 1k | 2k | 4k (default 2k) | A tier above the default costs more and needs the user to have asked for it |
quality | low | medium | high | xhigh | max (default high) | low garbles small UI text — raise only with user approval |
background | auto | transparent | opaque | Transparent is what game assets usually need |
prompt is ≤ 7500 chars and describes one finished result; count (1–9) is only for variations
of the SAME prompt. Up to 16 image references, ordered — the prompt can address “FIRST reference
image” / “LAST reference image”.
There are no pixel sizes. A request is a ratio plus a 1k/2k/4k tier (1:1 at 1k returned 1024×1024; 9:16 at 1k returned a ~576×1024-class portrait). To hit an exact game-asset size, generate the nearest ratio and crop/pad/resize locally. An aspect the model cannot express (a 4.5:1 slot when the widest ratio is 3:1) means keeping proportions and leaving transparent margins, or stretching — say which one you did.
Costs (observed, gpt-image-2.5-sunburst at 1k / quality low)
Section titled “Costs (observed, gpt-image-2.5-sunburst at 1k / quality low)”No references: 20 credits per image. 3 references: 50. 4: 60. 5: 70. References raise the price,
so import once and reuse. At ~60 credits/image no confirmation prompt appeared in the observing
session — the trigger threshold is UNVERIFIED, so never assume a batch is cheap. execute reports
billing.estimatedCredits per task.
Parallelism and waiting
Section titled “Parallelism and waiting”Independent execute calls may be sent together (13–15 at once was observed working); some return
status: "completed" immediately, the rest status: "running". For a running task, call
task_wait(taskId) with that same taskId until it completes — never resubmit, which starts and
bills new work. Queue waits of several minutes are normal at 13+ parallel tasks; one task sat
provider_queued for ~15 minutes and still completed.
Hard rules
Section titled “Hard rules”| Symptom | Cause | Fix |
|---|---|---|
PROJECT_NOT_OPEN | The project is not open in the MiniMax Design app | Ask the user to open it, retry open_project |
| Tools missing right after registering the server | The tool list is snapshotted at session start | Reconnect (/mcp) or restart the session |
INVALID_ARGUMENT … inside the workspace on deliver | The path is outside the project folder | Copy the file into a subfolder of the project, then deliver from there |
OUTPUT_UNAVAILABLE on get_output | A user-created version with no readable content (not on disk) | Ask the user to re-upload or save it into the project folder; never invent its content |
running after execute | Provider queue | task_wait the same taskId; never resubmit |
| Image is 1024×1024 for a 129×129 slot | The model has no pixel sizes | Resize/crop locally |
| Small UI labels come out wrong | quality: low | Raise the tier only with the user’s approval |
| Generated light rays / borders turn to noise at small sizes | Fine detail below the target resolution | Ask for “no sharp jagged rays” / a very thick border, regenerate that one image |
Image and media generation inside a MiniMax Design project, through the minimax-design MCP server.
Refuses: hardcoded project ids, ports, or credentials; writing into the user’s existing Figma
frames; and any spend the user has not agreed to. Project-specific conventions (game folders, art
style, naming) belong in project memory — do not add them here.
Lessons that outlive this page go to the gotcha ledger under skills/t1k-rsi/ledger/gotchas/ (as
t1k-minimax-design.md once one is recorded) via a [t1k:lesson] marker; a wrong step in this skill
itself is a [t1k:skill-bug].
Related
Section titled “Related”references/tool-reference.md— exact tool payloads and error codesreferences/production-playbook.md— smoke/wave batching, level progression, reskins, local resize, vision QA, Figma hand-offmodules/t1k-multimodal/rules/image-analysis-routing.md— why image review goes through the router vision roleskills/t1k-mcp-management/SKILL.md— installing, inspecting and guarding MCP servers