Happy Dreamina Skill for Claude

happy-dreamina

Generate images and videos through ByteDance's official dreamina CLI. This skill is a thin instruction layer — it does not wrap any SDK. Every action maps to a dreamina subcommand the user's shell already has.

When to invoke this skill

Use this skill whenever any of these hold:

• The user says 即梦, Jimeng, or Dreamina (any language).
• The user explicitly names the dreamina CLI or its subcommands.
• The user asks about their Jimeng task history, account credit, or login status.

If the user just says "generate an image" or "画张图" without naming Jimeng, prefer happy-image-gen / happy-video-gen instead and do NOT trigger this skill. The reason is that Jimeng uses browser login and is one specific provider — other skills let the user pick from many providers via EXTEND.md defaults.

Step 0: Preflight (BLOCKING — run both checks before anything else)

Run these in parallel:

• command -v dreamina — is the binary installed?
• dreamina user_credit — is the login alive? A healthy response is JSON containing credit info.

If dreamina is not installed

Tell the user the dreamina CLI is missing and offer to install it. The official installer is:

curl -fsSL https://jimeng.jianying.com/cli | bash

This is a shell install that writes to the user's machine — do not run it silently. Confirm, then run it, then re-check command -v dreamina. See references/install-and-login.md for platform notes (macOS, Linux x86_64/arm64). For any other platform (e.g. Windows) or when the installer changes, point the user at the official install page: https://jimeng.jianying.com/ai-tool/install.

If user_credit fails (exit non-zero, or JSON missing credit)

The user is not logged in, or the token expired. Ask them to run:

dreamina login

This opens the default browser for Jimeng authorization. Credentials land in ~/.dreamina_cli/credential.json automatically — the user never handles the file. If the browser does not open or the flow hangs, escalate to dreamina login --debug (see references/troubleshooting.md).

Do not proceed to generation until dreamina user_credit succeeds.

Step 1: Pick the right subcommand

Map the user's intent to exactly one subcommand:

| User wants to… | Command |

|---|---|

| Generate an image from text | dreamina text2image |

| Transform or restyle an existing image | dreamina image2image |

| Generate a video from text | dreamina text2video |

| Animate a still image (i2v) | dreamina image2video |

| Fetch an earlier async result | dreamina query_result |

| Browse past jobs | dreamina list_task |

Step 2: Fill in parameters from the user's intent

Use references/ratio-resolution-map.md to convert natural language ("竖屏 / 1080P / 高清 / 方图 / 横屏") to flags. Pay attention to what the user did not say — fill in safe defaults, do not ask the user to restate obvious things.

Image generation — text2image / image2image

• --prompt="..." — always double-quote so shell metacharacters (Chinese quotes, commas) do not break.
• --ratio — 1:1 / 16:9 / 9:16 / 3:4 / 4:3. Default 1:1.
• --resolution_type — 1k / 2k / 4k. Default 2k; use 1k for fast draft, 4k only when the user asks for ultra or print-ready.
• For image2image only: --images <path> (plural — accepts a reference bundle, multiple paths space-separated).
• Always include --poll=30.

Video generation — text2video / image2video

• --prompt="..." — same quoting rule.
• --duration — integer seconds. Default 5.
• --ratio — default 16:9 for landscape intent, 9:16 for vertical, 1:1 for square.
• --video_resolution — 480P / 720P / 1080P. Default 720P.
• For image2video only: --image <path> (singular — a single keyframe).
• Always include --poll=60 (videos take longer than images).

> ⚠ Easy to confuse: image2image uses --images (plural, reference bundle). image2video uses --image (singular, one keyframe). Mixing them up gives "unknown flag" errors.

Step 3: Submit and read the result

Run the command via Bash. On success the CLI prints JSON to stdout — the result URL(s) or local file path(s) if --download_dir was passed. Echo the path/URL back to the user.

Example — text to image:

dreamina text2image \
  --prompt="一只戴墨镜的橘猫,背光,电影感" \
  --ratio=1:1 \
  --resolution_type=2k \
  --poll=30

Example — image to video:

dreamina image2video \
  --image=./keyframe.png \
  --prompt="镜头缓缓推近,橘猫眼神闪动" \
  --duration=5 \
  --ratio=16:9 \
  --video_resolution=720P \
  --poll=60

If the user wants the result downloaded locally rather than a URL, add --download_dir=./out (create the directory first).

Step 4: Handle async timeout

If --poll times out, the CLI returns JSON with status querying and a submit_id. Tell the user you will retry, then:

dreamina query_result --submit_id=<submit_id> --download_dir=./out

If the job is still in progress, wait and retry. If it failed, pull the error message from the query_result response and relay it verbatim — do not guess.

Advanced: history and sessions

• dreamina list_task --gen_status=success — list completed jobs.
• dreamina list_task --submit_id=<id> — fetch one specific job.
• Sessions (v1.3.5+) let the user keep separate workspaces for different projects. Only touch them if the user explicitly asks to "switch session", "work in session X", or "list sessions". Otherwise the default session is fine.

References

Load these on demand — do not read all up front:

• references/cli-commands.md — full flag table for every dreamina subcommand.
• references/install-and-login.md — installer details, browser login, credential file location, version upgrade.
• references/ratio-resolution-map.md — user phrasing ↔ --ratio / --resolution_type / --video_resolution mapping.
• references/troubleshooting.md — login expired, queue timeout, submit_id recovery, account switching.