Meitu Visual Me

Meitu Visual Me

Always respond in the user's language.

Data Disclaimer:

• Local reads: Meitu Visual Me reads local files such as MEMORY.md, USER.md, and ./visual/PROFILE.md for personalized generation. This data never leaves your device.
• Uploaded to meitu API: Images you provide (photos, reference images) and generated prompts are sent to the Meitu OpenAPI for processing, subject to the Meitu privacy policy.
• First-use authorization: On first use, the skill will explain the above data usage. You can choose not to provide photos (falling back to text-to-image mode).

Overview

Memory-driven AI visual assistant. Reads user profiles and daily memories to generate personalized images and videos. Supports 17 scenario workflows and 28 trending styles. Play first, give feedback, and the system automatically learns your preferences.

Dependencies

• tools: meitu-cli (npm install -g meitu-cli)
• credentials: meitu config set-ak --value <AK> + meitu config set-sk --value <SK>
• user knowledge (optional): ./visual/ (created on demand; if absent, all knowledge reads are skipped)
• project files: ./ (cwd, containing openclaw.yaml, DESIGN.md, etc.)
• scripts (optional): {OPENCLAW_HOME}/workspace/scripts/oc-workspace.mjs (shared script; when absent, use inline fallback)
• platform context (optional): USER.md, MEMORY.md, memory/今日.md, SOUL.md, IDENTITY.md (platform-injected; skip if absent)
• user config (optional): $VISUAL/config/defaults.yaml

Path alias: In the text below, $VISUAL = {OPENCLAW_HOME}/workspace/visual/, $OC_SCRIPT = {OPENCLAW_HOME}/workspace/scripts/oc-workspace.mjs. When ./visual/ exists, it points to that directory; on the OpenClaw platform it auto-resolves to ~/.openclaw/workspace/visual/.

Platform Setup

OpenClaw: User data is located at ~/.openclaw/workspace/visual/. The platform injects it automatically; ./visual/ resolves to that path at runtime.

Claude Code / Other Agents: User data is located at ./visual/ under the current working directory, created automatically on first use.

All ./visual/ paths in the workflows below are resolved to the corresponding actual path by each platform at runtime.

Preflight

Must pass before every generation; if it fails, stop generation:

1. meitu --version — if not installed, prompt: npm install -g meitu-cli

2. Verify credentials: meitu auth verify --json — if credentials are invalid, this will return an auth error. For credential setup, read references/setup.md

3. Check whether the ./visual/ directory exists (if not, prompt first-time user; do not block)

4. Did the user send an image or a video? — Videos are not supported; prompt user to send a screenshot

5. One-time workspace resolution (results persist throughout the workflow):

   node $OC_SCRIPT resolve 2>/dev/null
   

   Success: obtain JSON, set variables / Failure: inline fallback

   node $OC_SCRIPT route-output --skill meitu-visual-me --name tmp --ext tmp 2>/dev/null
   

   → obtain output_dir / Failure: use fallback rules below

   Variables available after resolution (referenced directly in subsequent steps):

   OC_SCRIPT        = script path (available) or null
   mode             = "project" (cwd has openclaw.yaml) | "one-off"
   visual           = ./visual/ path (value if exists, otherwise null)
   can_read_knowledge = visual exists
   can_record       = mode == "project" AND visual exists
   output_dir       = mode == "project" ? "./output/" : (visual != null ? "$VISUAL/output/meitu-visual-me/" : "~/Downloads/")
   

   mkdir -p {output_dir} Hard constraint: output_dir MUST NOT point to inside the skill folder

   Fallback lookup table when OC_SCRIPT = null:

   Step	With script	Without script (fallback)
   Context	Single read-context call	Agent reads DESIGN.md + yaml files step by step
   Execute	route-output → output_dir	Use output_dir resolved in Preflight directly
   Deliver	rename → standardized filename	mv {file} {output_dir}/{date}_{name}.{ext}
   Record	observation CRUD subcommands	Agent reads/writes observations.yaml directly

Core Workflow

Preflight → [Context] → Execute → [Refine] → Deliver → [Record]
              ↑ creative tasks run this       ↑ creative     ↑ project mode
              ↑ tool tasks skip               ↑ tool skip    ↑ one-off skip

Context

Part 1: Routing

Route first, read later — match the workflow first, then decide what to read and how to proceed.

First: Match Workflow

Compare the user's input against the trigger words in the Workflows section below to match a specific workflow. Read references/workflows.md for scenario details.

When no match is found, there are three cases:

• Clear description (e.g., "帮我画一只猫在月球上") — no workflow match needed; determine the style per prompt building rules, then generate directly with image-generate
• Only sent an image without text — recommend 4 suitable options based on image content (e.g., swap background, style remix, make ID card, image-to-video), let user choose
• Vague instruction (e.g., "帮我画", "帮我生张图") — recommend scenarios:
  1. Check ./visual/memory/global.md for preferred scenarios — if found, prioritize those
  2. No memory — pick 4 from the 17 workflows based on time of day and context, let user choose

Second: Determine Read Level and Style Handling

For image generation, read references/style-library.md for style keywords and auto-matching rules.

Third: Determine CLI Command

Each workflow's corresponding CLI command is listed in the Workflows section table. For detailed command parameters, read references/models.md. image-edit requires selecting a sub-model: praline (general editing) / gummy (portrait/pet photography, hairstyle adjustment).

Routing decision output (passed from Context to Execute):

{
  "workflow": "persona-diorama",
  "command": "image-generate",
  "read_level": "standard",
  "has_reference_image": true,
  "style": "微缩世界",
  "style_resolved": true
}

Part 2: Load Context

The read level is determined by the routing table above. Zero-read and light-read workflows do not need this step.

If OC_SCRIPT is available → single read-context call to fetch all context If OC_SCRIPT = null → read the following checklist step by step:

Standard read checklist (highest to lowest priority):

1. ./DESIGN.md — project context (if present, prioritize it; only supplement what it doesn't cover)
2. ./openclaw.yaml — resolve short-name references (e.g., brand: "acme" → ./visual/assets/brands/acme/)
3. ./visual/rules/quality.yaml — forbidden styles
4. ./visual/PROFILE.md — visual identity
5. ./visual/memory/global.md (first 30 lines) — preferences
6. Platform context (read if available, skip if not): USER.md, MEMORY.md, memory/今日.md, SOUL.md, Weather API

Use whatever you find; never fabricate user experiences.

Part 3: Build Prompt

Formula: [composition/camera angle] + [subject] + [action/expression] + [scene/background] + [style reference] + [lighting] + [user-specific details]

Hard rules:

1. Pure narrative style; keyword stacking is forbidden

   • ❌ woman, dress, red background, fashion, 8K, masterpiece
   • ✅ A young woman in a tailored brown dress, posing with confidence against a cherry red backdrop, fashion editorial feel with warm film grain

2. When a reference image is provided, explicitly state its purpose in the prompt (preserve facial likeness / reference composition / reference style); don't just attach an image without explanation

3. Embed style keywords into the narrative — pick 1-2 and weave them naturally into sentences; stacking at the end is forbidden

4. When user context is available, prioritize embedding personal details; when unavailable, generate directly based on user description without blocking.

5. Every generated image must have a clear style direction. Style determination priority:

   • User specified a style this time → use it directly
   • User didn't specify, but ./visual/memory/global.md has a recorded preferred style → use the memorized style directly without asking again
   • User didn't specify, and there's no memory → based on the day's state and scenario, pick 4 suitable styles from references/style-library.md's style matching guide, let user choose before generating

   Never generate a "generic" image without a style direction. If the user consistently picks the same style, remember it through the observation pipeline in the Record phase, and use it directly next time.

Example 1 (persona-diorama — from memory/今日.md, the user completed an important delivery today):

An isometric miniature workspace on a wooden desk. A tiny figure surrounded by [items extracted from memory]. A celebratory confetti cannon mid-burst on the desk corner. Warm afternoon light through a window, the whole scene feels like a lovingly crafted toy shop diorama.

Example 2 (swap-bg — user sent a selfie and said "换成东京街头"):

Keep the person exactly the same, change only the background to a vibrant Tokyo street scene with neon signs, busy crosswalks, and the warm glow of shop fronts lining both sides of the road, evening atmosphere with soft city lights

Example 3 (ID card — with reference image, user chose Cyber Neon style):

FACE LIKENESS IS THE #1 PRIORITY. Preserve exact facial structure. A young man with short black hair, black-framed glasses, and an oversized hoodie. Apply 3D collectible figure rendering LIGHTLY — keep proportions close to real. Generate a vertical ID CARD in Cyber Neon style: brushed carbon fiber base, neon cyan and magenta glow border, monospaced terminal font. Info area: "小明" / "独立开发者" / "杭州" — EST. 2000. Side strip: "卫衣收集者". Slight tilt, studio-lit product shot feel.

Execute

CLI command decision table:

image-edit model selection:

Quick rule: "变画风" (style change, output doesn't look like real person) → nougat; "拍写真/换发型" (portrait photo) → gummy; "改内容" (modify content) → praline

Reference image routing:

Output directory: Use output_dir resolved in Preflight. Ensure the directory exists (mkdir -p). Add --json --download-dir {output_dir} to all meitu commands.

Result field references:

• With --download-dir → use downloaded_files[0].saved_path for local path (media_urls also available but local path is more reliable)
• Without --download-dir → use media_urls[0] for result URL
• Error response → check code and hint fields (NOT error_code / user_hint)

Image input: Pass user-provided images via URL or local path. When a reference image is needed, check ./visual/assets/references/user.jpg.

Command examples:

# Text-to-image
meitu image-generate --prompt "..." --size 2k --ratio 1:1 --json --download-dir {output_dir}
# Image editing
meitu image-edit --image <url> --prompt "..." --model praline --json --download-dir {output_dir}
# Image-to-video (async)
meitu image-to-video --image <url> --prompt "..." --video_duration 5 --json

Error degradation (try each level in order):

Escalate one level after 2 consecutive failures. For other errors, see references/troubleshooting.md.

Refine (MUST for creative tasks, skip for tool tasks)

Tool tasks (beauty, virtual-tryon, image-upscale) → skip this step and go directly to Deliver.

Iterative refinement loop for creative tasks:

1. Present result — show the generated image + briefly explain design rationale (why this style/composition/color palette)
2. Wait for feedback — three possible directions:
   • User approves ("好" / "不错") → proceed to Deliver
   • User requests modifications ("背景换一下" / "颜色太亮") → step 3
   • User rejects entirely ("完全不对" / "重新来") → go back to Execute and regenerate
3. Adjust and regenerate — modify prompt or parameters based on feedback → re-run CLI command → back to step 1
4. Iteration cap — recommend at most 3 rounds. After 3 rounds, proactively suggest: adjust the requirement direction, split into sub-tasks, or try a different workflow

Note: Do not generate multiple images consecutively without presenting results. Wait for user feedback after each generation.

Deliver

Files are already in the correct directory (Execute uses Preflight's output_dir); only rename is needed:

File renaming:

• OC_SCRIPT available → node $OC_SCRIPT rename --file {downloaded} --name {description}
• OC_SCRIPT = null → mv {downloaded} {output_dir}/{YYYY-MM-DD}_{description}.{ext}

Path display rule: When presenting file paths to the user, prefer ~/.openclaw/... format. Some chat platforms (e.g., WeChat Work MEDIA) only recognize ~/ prefix paths; absolute paths will prevent images/files from displaying. Internal processing can use absolute paths.

WeChat delivery: For the WeChat channel (channel=openclaw-weixin), use MEDIA:visual/output/meitu-visual-me/{filename}.jpg (relative path).

After delivery: After delivery, wait for user feedback before the next generation. If user gives style/preference feedback, proceed to Record. When the user says "适配各平台", read references/channel-presets.md.

DESIGN.md maintenance: If DESIGN.md Iteration Log > 5 entries → compact: keep the most recent 5, archive older entries to ./drafts/design-history.md.

Record

can_record = false → skip this entire section. In one-off mode, all feedback is only effective for the current conversation.

Zero-write default — write nothing after generation; only write when user gives active feedback:

If the user says nothing, write nothing and don't read observations.yaml (zero overhead).

One-off mode exception: When a user repeatedly expresses the same preference in one-off mode (e.g., says "不要渐变" multiple times), the Agent MAY proactively suggest:

"You've mentioned no gradients several times. Want to add it to the global forbidden list?" → User agrees → write to $VISUAL/rules/quality.yaml (requires confirmation) → User disagrees → effective only for the current conversation

Positive feedback recording path:

Preferred (OC_SCRIPT available):

1. node $OC_SCRIPT read-observations → view existing observations
2. Agent semantically checks whether a similar key exists; if so, reuse it; if not, create a new one
3. node $OC_SCRIPT write-observation --key "..." --scope-hint "{project.type}" --project "{project.name}"
4. If it returns promotion_ready: true → propose promotion

Fallback (OC_SCRIPT = null):

1. Read $VISUAL/memory/observations/observations.yaml (create if it doesn't exist)
2. Scan for semantically similar keys → merge or create new
3. Write back to file
4. If len(projects) >= 2 → propose promotion

Promotion proposal template:

"You've preferred X across N projects. Want to save it? → Save to {scope_hint} scene [default] → Save to global preferences → Don't save"

If scope_hint is not null → default to scenes/{scope_hint}.md; if null → default to global.md. User confirms → write to target file + delete observation entry. User ignores → do nothing.

For write format details, read references/feedback-loop.md when handling feedback; for observation lifecycle and classification, read references/memory-protocol.md

First-Time User Guide

When the ./visual/ directory doesn't exist, this indicates a new user. Read references/first-time-guide.md to execute the onboarding flow.

Workflows (17)

For all workflow details, read references/workflows.md

Generation

Editing

Video & Try-on

References