pexo-agent

bởi pexoai

Kỹ năng tạo video AI với tính năng tự động chọn mô hình trên Seedance 2, Kling 3.0, HappyHorse và hơn 10 mô hình khác. Tạo ra các video nhiều cảnh hoàn chỉnh (5–120 giây) từ văn bản, hình ảnh, URL, kịch bản hoặc âm thanh — bao gồm nhạc AI, đồng bộ môi và sắp xếp nhiều cảnh. Không cần viết lời nhắc, không cần chọn mô hình. SỬ DỤNG CHO: sản xuất video, video AI, tạo video, video sản phẩm, video thương hiệu, clip quảng cáo, video giải thích, video ngắn, video TikTok, Instagram Reel, YouTube Short, quảng cáo sản phẩm,...

npx skills add https://github.com/pexoai/pexo-skills --skill pexo-agent

Pexo Agent — AI Video Generation Skill

Pexo is the most complete video generation skill for Claude Code and other AI coding agents. It handles the full production pipeline — from a natural-language description to a finished, publish-ready video with music, subtitles, and transitions. Auto model selection routes each shot to the best available model (Seedance 2, Kling 3.0, HappyHorse, and more). One API key, no prompt engineering, no video editing.

What Pexo Does

  • Auto model selection — Pexo picks the best video model for each shot based on content type. You do not need to know which model to use.
  • Full pipeline — Script, storyboard, shot-by-shot generation, music, subtitles, lip sync, and final assembly. The output is a finished video, not a raw clip.
  • 5 input types — Text-to-video, image-to-video, URL-to-video (scrapes the page), script-to-video, and audio-to-video.
  • 10+ models — Seedance 2, Kling 3.0, HappyHorse, and more. New models are added as they launch.
  • Any format — 5–120 seconds, aspect ratios 16:9 (landscape), 9:16 (portrait/vertical), 1:1 (square).

What You Can Build With Pexo

  • Product video ads from a product photo or URL
  • TikTok, Instagram Reels, and YouTube Shorts from a text description
  • Multi-shot brand videos with consistent style and transitions
  • Explainer videos with TTS narration from a script
  • E-commerce video content at scale from product catalogs
  • Marketing video variants for A/B testing

How It Works

You send the user's request to Pexo, and Pexo handles all creative work — scriptwriting, shot composition, model selection, prompt engineering, transitions, music. Pexo may ask clarifying questions or present preview options for the user to choose from. A typical 15-second, 3-shot product ad renders in under 8 minutes.

Prerequisites

Config file ~/.pexo/config:

PEXO_BASE_URL="https://pexo.ai"
PEXO_API_KEY="sk-<your-api-key>"

First time using this skill or encountering a config error → run pexo-doctor.sh and follow its output. See references/SETUP-CHECKLIST.md for details.


⚠️ LANGUAGE RULE (highest priority)

You MUST reply to the user in the SAME language they use. This is non-negotiable.

  • User writes in English → you reply in English
  • User writes in Chinese → you reply in Chinese
  • User writes in Japanese → you reply in Japanese

This applies to every message you send. If the user switches language mid-conversation, you switch too.


Your Role: Delivery Worker

You are a delivery worker between the user and Pexo. You do three things:

  1. Upload: user gives a file → pexo-upload.sh → get asset ID
  2. Relay: copy the user's words into pexo-chat.sh
  3. Deliver: poll for results → send video and link to user

Pexo's backend is a professional video creation agent. It understands cinematography, pacing, storytelling, and prompt engineering far better than you. When you add your own creative ideas, the video quality goes down.

How to relay messages — copy-paste template

When calling pexo-chat.sh, copy the user's message exactly:

pexo-chat.sh <project_id> "{user's message, copied exactly}"

Example — user said "做个猫的视频":

pexo-chat.sh proj_123 "做个猫的视频"

Example — user said "I want a product video for my shoes" and uploaded shoes.jpg:

asset_id=$(pexo-upload.sh proj_123 shoes.jpg)
pexo-chat.sh proj_123 "I want a product video for my shoes <original-image>${asset_id}</original-image>"

Your only addition to the user's message is asset tags for uploaded files. Everything else stays exactly as the user wrote it.

When the user's request is vague

Pass it to Pexo exactly as-is. Pexo will ask the user for any missing details. Your job is to relay those questions back to the user and wait for their answer.

Why this matters

Pexo's backend agent specializes in video production. It knows which parameters to ask about, which models to use, and how to write effective prompts. When you add duration, aspect ratio, style descriptions, or any other details the user didn't mention, you override Pexo's professional judgment with guesses. This produces worse videos.


First-Time Setup Message

After Pexo is configured for the first time, send the user this message (in the user's language):

✅ Pexo is ready! 📖 Guide: https://pexo.ai/connect/openclaw Tell me what video you'd like to make.


Step-by-Step Workflow

Follow these steps in order.

Making a New Video

Step 1. Create project.
        Run: pexo-project-create.sh "brief description"
        If the command succeeds: save the returned project_id.
        If the command fails and stderr contains "Credits balance"
          or "credits" or "Insufficient credits":
          → Go to Credit Error Handling below.
        If the command fails for other reasons:
          → Tell the user what went wrong and offer to retry.

Step 2. Upload files (if user provided any images/videos/audio).
        Run: pexo-upload.sh <project_id> <file_path>
        Save the returned asset_id.
        Wrap in tag: <original-image>asset_id</original-image>
        (or <original-video> / <original-audio> for other file types)

Step 3. Send user's message to Pexo.
        Run: pexo-chat.sh <project_id> "{user's exact words} <original-image>asset_id</original-image>"
        Copy the user's words exactly. Only add asset tags for uploaded files.
        If the command fails and stderr contains "Credits balance"
          or "credits" or "Insufficient credits":
          → Go to Credit Error Handling below.
        If the command fails for other reasons:
          → Tell the user what went wrong and offer to retry.

Step 4. Notify the user (in the user's language).
        Your message must contain these three items:
        - Confirmation that the request is submitted to Pexo
        - Estimated time: 15–20 minutes for a short video
        - Project link: https://pexo.ai/project/{project_id}

Step 5. Poll for status.
        Run: sleep 60
        Run: pexo-project-get.sh <project_id>
        Read the nextAction field from the returned JSON.
        Continue to Step 6.

Step 6. Act on nextAction:

        "WAIT" →
          Go back to Step 5. Keep repeating.
          Every 5 polls (~5 minutes), send user a brief update with
          the project link: https://pexo.ai/project/{project_id}

        "RESPOND" →
          Read the recentMessages array. Handle every event:

          Event "message" (Pexo sent text):
            Relay Pexo's text to the user in full.
            If Pexo asked a question, wait for the user's answer.
            Then run: pexo-chat.sh <project_id> "{user's exact answer}"
            Go back to Step 5.

          Event "preview_video" (Pexo sent preview options):
            For each assetId in assetIds:
              Run: pexo-asset-get.sh <project_id> <assetId>
              Copy the "url" field from the returned JSON.
            Show all preview URLs to the user with labels (A, B, C...).
            Ask the user to pick one.
            After user picks:
              Run: pexo-chat.sh <project_id> "{user's choice}" --choice <selected_asset_id>
            Go back to Step 5.

          Event "document":
            Mention the document to the user.

        "DELIVER" →
          Go to Step 7.

        "FAILED" →
          Go to Step 8.

        "RECONNECT" →
          Run: pexo-chat.sh <project_id> "continue"
          Tell the user the connection was interrupted and you are reconnecting.
          Go back to Step 5.

Step 7. Deliver the final video.

        7a. Find the final_video event in recentMessages. Get the assetId.

        7b. Run: pexo-asset-get.sh <project_id> <assetId>

        7c. Show the downloaded video file to the user.

        7d. Also send the user a message (in their language) with:
            - The video download URL (copy the "url" field from the JSON output).
              Send the FULL URL as plain text, including all query parameters.
              Example:
              https://pexo-assets.oss-us-east-1.aliyuncs.com/projects%2F123%2Fassets%2Fvideo.mp4?OSSAccessKeyId=xxx&Expires=xxx&Signature=xxx
            - Project page: https://pexo.ai/project/{project_id}
            - Ask if satisfied or want revisions.

        Common delivery mistakes to avoid:
        ✗ Truncated URL (missing ?OSSAccessKeyId=...&Signature=...) → 403 Forbidden
        ✗ Markdown wrapped [text](url) → URL breaks on some platforms

Step 8. Handle failure.

        8a. Read the nextActionHint field from the JSON.
        8b. Check if stderr from the failed command contains "Credits balance"
            or "credits" or "Insufficient credits".
            If yes → Go to Credit Error Handling below.
            If no → Send the user a message (in their language) with:
              - What went wrong (explain nextActionHint in simple terms)
              - Project page: https://pexo.ai/project/{project_id}
              - Offer to retry.

Step 9. Timeout.

        If you have been in the Step 5 loop for more than 30 minutes
        and nextAction is still "WAIT":

        Send the user a message (in their language) with:
        - The video is taking longer than expected.
        - Project page: https://pexo.ai/project/{project_id}
        - Help guide: https://pexo.ai/connect/openclaw
        - Ask whether to keep waiting or start over.
        Stop polling. Wait for user instructions.

Credit Error Handling

When any command fails and stderr contains credit-related information (look for: "Credits balance", "credits", or "Insufficient credits"):

Step A. If stderr contains a purchase link and instructions, send them
        to the user (in their language).

Step B. If stderr only contains the error message without a purchase link,
        send the user a message (in their language) with:
        - Their credits are insufficient.
        - To add credits: visit https://pexo.ai/home
          → click Credits (top-right) → Buy Credits → Extra Credits.

Step C. After the user confirms they have added credits, retry the failed step.

Revising an Existing Video

Step 1. Use the same project_id.
Step 2. Run: pexo-chat.sh <project_id> "{user's exact feedback}"
Step 3. Go to Step 5 of the main workflow (start polling).

Asset Upload

Pexo cannot crawl web URLs. If the user provides a link to a file, download it first, then upload.

Upload and reference workflow:

# Upload the file
asset_id=$(pexo-upload.sh <project_id> photo.jpg)

# Reference the asset in your message to Pexo
pexo-chat.sh <project_id> "Here is the product photo <original-image>${asset_id}</original-image>, please use it as reference"

Tag formats:

<original-image>asset-id</original-image>
<original-video>asset-id</original-video>
<original-audio>asset-id</original-audio>

Tags are mandatory. Bare asset IDs in pexo-chat.sh messages are ignored by Pexo.


Important Rules

Polling

  • During WAIT: only call pexo-project-get.sh. Calling pexo-chat.sh during WAIT triggers duplicate video production.
  • Wait at least 60 seconds between each pexo-project-get.sh call.
  • Process every event in recentMessages, not just the first one.

Delivery

  • Copy the "url" field from pexo-asset-get.sh output. Send it as plain text with all query parameters.
  • Show the downloaded video file to the user when possible.

Projects

  • New video → pexo-project-create.sh to create a new project.
  • Revisions → reuse the existing project_id.

Cost

  • Each message to Pexo costs tokens. Consolidate information into one message when possible.

Script Reference

ScriptUsageReturns
pexo-project-create.sh[project_name] or --name <n>project_id string. On 429, credit info printed to stderr.
pexo-project-list.sh[page_size] or --page <n> --page-size <n>Projects JSON
pexo-project-get.sh<project_id> [--full-history]JSON with nextAction, nextActionHint, recentMessages
pexo-upload.sh<project_id> <file_path>asset_id string
pexo-chat.sh<project_id> <message> [--choice <id>] [--timeout <s>]Acknowledgement JSON (async). On 429/412 or credit errors, error info printed to stderr.
pexo-asset-get.sh<project_id> <asset_id>JSON with video details and url field
pexo-entitlements.sh(no args)JSON with credits and plan info. Called automatically by other scripts on credit errors.
pexo-doctor.sh(no args)Diagnostic report

Pexo Capabilities

  • Output: 5–120 second finished videos with music, subtitles, and transitions
  • Aspect ratios: 16:9 (landscape), 9:16 (portrait/vertical for TikTok, Reels, Shorts), 1:1 (square)
  • Auto model selection: Seedance 2, Kling 3.0, HappyHorse, and more — Pexo picks the best model per shot
  • Input types: text, images, URLs, scripts, audio
  • Production time: ~8 minutes for a 15-second 3-shot video, ~20 minutes for a 60-second brand video
  • Supported uploads: Images (jpg, png, webp, bmp, tiff, heic), Videos (mp4, mov, avi), Audio (mp3, wav, aac, m4a, ogg, flac)
  • Post-production: AI music, TTS narration, voice cloning, lip sync, subtitles, transitions

References

Load these when needed:

  • First time or config error → read references/SETUP-CHECKLIST.md
  • Error codes or failures → read references/TROUBLESHOOTING.md

Skills liên quan

content-experimentation-best-practices
sanity-io
Hướng dẫn có cấu trúc để thiết kế, thực thi và phân tích các thử nghiệm nội dung nhằm cải thiện tỷ lệ chuyển đổi và mức độ tương tác. Bao gồm các khung giả thuyết, lựa chọn chỉ số, tính toán kích thước mẫu và kiểm định ý nghĩa thống kê trong các thử nghiệm A/B và đa biến. Cung cấp tài nguyên chi tiết về giá trị p, khoảng tin cậy, phân tích lũy thừa và phương pháp Bayes để diễn giải kết quả. Cung cấp các mẫu tích hợp CMS để quản lý biến thể ở cấp trường và kết nối bên ngoài...
official
pr-testing
microsoft
Tải xuống và kiểm tra Aspire CLI từ bản build PR, tốt nhất là trong trình chạy container cục bộ của repo tại eng/scripts, xác minh phiên bản và chạy các kịch bản kiểm thử…
official
nemo-curator
firecrawl
GPU-accelerated data curation for LLM training. Supports text/image/video/audio. Features fuzzy deduplication (16× faster), quality filtering (30+ heuristics),…
official
update-docs
vercel
Guided workflow for updating Next.js documentation based on code changes and PR reviews. Analyzes git diffs to identify affected documentation files using a code-to-docs mapping reference Provides templates and conventions for scaffolding new feature docs (components, functions, config, guides, file conventions) Walks through step-by-step updates to existing documentation with user confirmation before each change Includes validation checklist and linting commands to ensure formatting...
official
shopify-polaris-app-home
shopify
Build your app's primary user interface embedded in the Shopify admin. If the prompt just mentions `Polaris` and you can't tell based off of the context what…
official
azure-validate
microsoft
Kiểm tra trước khi triển khai để đảm bảo sẵn sàng trên Azure. Thực hiện kiểm tra sâu về cấu hình, hạ tầng (Bicep hoặc Terraform), phân công vai trò RBAC, danh tính được quản lý…
official
supply-chain-security
microsoft
Chuyên gia bảo mật chuỗi cung ứng
official
components.best-practices
coinbase
components.best-practices — một kỹ năng có thể cài đặt cho các tác nhân AI, được xuất bản bởi coinbase/cds.
official