text-generate

Text Generate

Overview

Independent LLM deep creation tool that generates pure text content such as copy/scripts/storyboards/video prompts from scratch. Covers marketing copy, product descriptions, social media copy, articles/blogs, advertising slogans, content polishing and rewriting, multilingual copy, image/video reference copy, video storyboard planning, and video prompt optimization. ⚠️ When the context already contains complete content, calling is prohibited; use write_file to write directly to disk instead; use only when the context contains only directional requirements and the LLM needs to think deeply independently. Does not handle: video/image/audio/code/table generation.

API Mapping

Scenario	Backend API	Underlying Model
General text / storyboard script generation	text_generate	qwen3.6-vl-plus（scone）/ gemini-3.1-pro-preview（pretzel_pro）

model mapping:

• auto(default) / not passed → decided by routing rules
• scone → qwen3.6-vl-plus(daily workhorse)
• pretzel_pro → gemini-3.1-pro-preview(strongest video understanding, long context)
• The two models serve as each other's fallback

Dependencies

• meitu-cli: >=2.0.6
• Credentials：CONFIG AKSK → meitu tools update; EXEC AKSK → run command (see root CONFIG.md）
• Environment Variables：MEITU_OPENAPI_TOOL_TASK_MODE=command

Core Workflow

Preflight → Execute → Deliver

Preflight

1. meitu --version ≥ 2.0.6
2. meitu auth verify --json
3. Confirm the registry contains text-generate
4. Resolve output_dir:openclaw.yaml → ./output/；else → ~/.openclaw/workspace/visual/output/text-generate/

Execute

Trigger Signals / Routing Rules

Scenario	Decision Dimension	Routing
Has video reference	video_list non-empty	Lock gemini-3.1-pro-preview（pretzel_pro）
High-difficulty scenario	Caller explicitly passes model=pretzel_pro	gemini-3.1-pro-preview
Daily scenario	auto / model not passed	qwen3.6-vl-plus(scone, daily default)

High-difficulty criteria (judged by the caller): video input, long context >2000 characters, complex reasoning (multi-step logic/comparative analysis/causal argumentation), professional domains (law/medicine/finance/academia), multi-constraint creation (≥3 creative constraints simultaneously), deep cross-language adaptation, multi-shot storyboard script planning.

Parameter Definitions

Parameter	Type	Required	Default Value	Description
summary	STRING	Yes	--	Copy/storyboard requirement description. Should include copy type, target audience, core objective, and style; storyboard scenarios need to include theme, number of shots, and scene description. If missing → ask follow-up questions
language	STRING	Yes	Simplified Chinese	Output language, consistent with the user's input language
model	STRING	No	According to routing rules	Optional scone / pretzel_pro
prompt_name	STRING	No	auto	Prompt optimization template name
image_list	ARRAY	No	--	List of reference image URLs; do not pass if none
video_list	ARRAY	No	--	List of reference video URLs; do not pass if none
output_file	STRING	No	--	Output file URL address

Tool Invocation

Daily copy:

meitu text-generate \
  --skill_name skill_text-generate \
  --summary "{需求描述}" \
  --language "简体中文" \
  --json

High-difficulty / video storyboard:

meitu text-generate \
  --skill_name skill_text-generate \
  --summary "{分镜需求}" \
  --language "简体中文" \
  --model pretzel_pro \
  --video_list "{video_url}" \
  --json

English copy:

meitu text-generate \
  --skill_name skill_text-generate \
  --summary "Write a product description for ..." \
  --language "English" \
  --json

Error Degradation

Scenario	Handling Method
summary missing	Prompt "Please describe what you want"
summary < 5 characters	Prompt to supplement copy type/target audience/style
language Not specified	Automatically determined by the user's input language, defaults to Simplified Chinese
image_list All invalid	Silently filter, select model via the plain text path
Default model call failed	Switch to another model and retry
Both models failed	Return error message
API timeout	Switch to another model and retry → if both fail, return error
Content compliance interception	Directly return compliance prompt

Conflict rules:

• User requests both storyboard + direct video output → this tool only generates storyboard scripts; video generation is chained by the upper-layer workflow text-to-video
• User requests copy + accompanying images → copy goes through this tool; accompanying images require upper-layer orchestration text-to-image
• User specifies a model name → respect the specification, skip automatic routing

Deliver

Parse --json：

• ok: true → extract text content from the returned structure
• If explicitly specified during the call output_file, then write according to that path
• Otherwise write the text to {output_dir}/{YYYY-MM-DD}_{descriptive-name}.md

Default persistence:{output_dir}/{YYYY-MM-DD}_{descriptive-name}.md

Output

• Format: plain text / Markdown (.md / .txt）
• Naming: {YYYY-MM-DD}_{descriptive-name}.md
• Location: project → ./output/; one-time → ~/.openclaw/workspace/visual/output/text-generate/

Baseline Task ID

See root directory references/task-id-baseline.md in text-generate entry.