# Tensorart Generate

> use TensorArt/Tusi/吐司 to generate image or video for you

- Skill: `tensor-art/tensorart-generate` (Agent Skill, multi-file: 7 files)
- Install (CLI): `npx skillmds@latest add tensor-art/tensorart-generate`
- Raw SKILL.md: https://api.skillmd.com/api/skills/tensor-art/tensorart-generate/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: tensor-art (https://skillmd.com/u/tensor-art)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/tensor-art/tensorart-generate

---


# TensorArt Image/Video Generation Skill

You are an image/video generation assistant. Help users generate images or videos via the TensorArt/Tusi/吐司 OpenAPI.

All API calls are made through Python scripts in the `scripts/` directory, located at: `~/.claude/skills/tensorart-generate/scripts/`.

**Always `cd` into the skill directory before running any script:**
```bash
cd ~/.claude/skills/tensorart-generate && python3 scripts/xxx.py ...
```

## User Request

$ARGUMENTS

## Step 1: Check Access Key

The Access Key is stored in `~/.tensor_access_key`. All scripts read this file automatically.

If a script reports `~/.tensor_access_key not found`, **stop and tell the user:**

> You haven't configured your TensorArt Access Key yet. Get one at:
> https://tensor.art/settings/access-key
>
> Then save it by running:
> ```
> echo "your-access-key" > ~/.tensor_access_key
> ```
> After that, re-run the generation command.

## Step 2: List Available Tools

```bash
cd ~/.claude/skills/tensorart-generate && python3 scripts/list_tools.py
```

Returns a full JSON list of all available tools, each with `name`, `description`, `inputs`, `outputs`, `estimatedCost`, and `tags`.

**Recommend the 3 best-matching tools for the user to choose from:**
- Analyze each tool's `name`, `description`, and `tags` against the user's intent
- Show the user a brief summary of each: tool name, description, estimated compute cost (`estimatedCost`), and use cases
- **Wait for the user to choose** before proceeding — do not decide automatically
- Remember the selected tool's `inputs` definition for use in the next steps

## Step 3: Prepare Inputs (if file upload is needed)

If the selected tool has any `type: FILE` inputs, you need a file URL. There are two cases:

### Case A: User provides a local file path

Upload directly (see 3.2).

### Case B: Using a previous task's output as input

Output URLs from previous tasks are signed temporary URLs and **cannot be used directly as FILE inputs**. You must download them locally first, then re-upload:

```bash
cd ~/.claude/skills/tensorart-generate && python3 scripts/download_result.py "${PREVIOUS_OUTPUT_URL}" /tmp/previous_result.png
```

Then upload the downloaded file as a local file (see 3.2).

### 3.2 Upload a File

```bash
cd ~/.claude/skills/tensorart-generate && python3 scripts/upload_file.py /path/to/local/file.png
```

The script automatically fetches an upload URL and PUTs the file to Cloudflare.

Output JSON: `{"displayUrl": "...", "accessUrl": "..."}`

Use **`displayUrl`** (if non-empty) or `accessUrl` as the file value in the task inputs. `displayUrl` is a stable URL that won't expire.

## Step 4: Create a Generation Task

```bash
cd ~/.claude/skills/tensorart-generate && python3 scripts/create_task.py "toolName" '[{"type":"STRING","value":"your prompt"}, ...]'
```

The second argument is a JSON array string. **Each element corresponds to the input at the same position in the tool definition:**
- `type`: one of `STRING`, `INTEGER`, `NUMBER`, `BOOLEAN`, `ARRAY`, `OBJECT`, `FILE`
- `value`: the value matching the type
- For `FILE`: use the `displayUrl` or `accessUrl` from Step 3
- For `OBJECT`: use a JSON object
- For `ARRAY`: use a JSON array

**All inputs are required:**
- Every input defined by the tool must have a **meaningful value**
- **Never** use placeholder values (e.g. `0`, `""`, `null`, `" "`)
- For dimensions (width/height): choose reasonable values (e.g. 512–1024 for images, 480–720 for video)
- For count: default to `1`
- For prompt/description fields: generate specific text based on the user's intent
- If unsure what value to use, infer a reasonable default from the input's `description`

Output JSON: `{"taskId": "...", "status": "..."}`

Record the `taskId` for the next step.

## Step 5: Poll Task Status

```bash
cd ~/.claude/skills/tensorart-generate && python3 scripts/query_task.py "${TASK_ID}" --poll
```

`--poll` mode auto-polls every 3 seconds (up to 60 attempts), printing status to stderr and outputting the final result JSON to stdout.

**Task status reference:**

| Status | Meaning |
|--------|---------|
| `WAITING` / `QUEUE_WAIT` / `PARSING` / `START` / `PROCESSING` | In progress |
| `FINISH` | Completed |
| `EXCEPTION` | Failed — report the error reason to the user |
| `CANCELED` | Canceled |

## Step 6: Show Results

When the task completes (`status: FINISH`):
1. Extract results from the `outputs` field of the response JSON
2. For `FILE` outputs, display the image/video URL using markdown syntax: `![result](url)`
3. Show all outputs if there are multiple
4. Inform the user of the estimated compute cost consumed

To save a result locally:
```bash
cd ~/.claude/skills/tensorart-generate && python3 scripts/download_result.py "${RESULT_URL}" /tmp/result.png
```

## Notes

- If the user's description is brief, you may enrich the prompt — but tell the user what you changed
- If the user writes in Chinese, consider translating the prompt to English (most models perform better with English prompts); keep the Chinese version visible to the user
- On errors, display the full error message to help with debugging