Giggle Generation Video

Giggle Video Generation

Source: giggle-official/skills · API: giggle.pro

Generates AI videos via giggle.pro's Generation API. Supports text-to-video and image-to-video. Submit task → agent proactively polls with --query until done (see Continuous progress updates). No Cron, no file writes—all operations via exec.

Installation Requirements

Where to get GIGGLE_API_KEY: Open giggle.pro while logged in → left sidebar → API Key (API 密钥) → create or copy your key. Then export GIGGLE_API_KEY=your_api_key. The script will prompt if not configured.

No Retry on Error: If script execution encounters an error, do not retry. Report the error to the user directly and stop.

Supported Models

Note: --duration must be chosen from the model's supported durations, otherwise the API will error.

Frame Reference (Image-to-Video)

For image-to-video, --start-frame and --end-frame support three mutually exclusive formats:

Each frame parameter can only use one of these methods.

Execution Flow: Submit and Query

Video generation is asynchronous (typically 60–300 seconds). Submit a task to get task_id, then query until the task reaches a terminal state. All commands run via exec; API key from system env.

Continuous progress updates (default; user need not put this in their prompt)

Video jobs are usually ~1–5+ minutes (sometimes longer). The user does not need to ask you to check progress.

1. Right after submit, say you submitted, give task_id, and give a realistic wait range for the model/load (not a generic “1–3 minutes” if the pipeline is often slower).
2. Poll proactively: run --query about every 15–30 seconds until a terminal state—do not wait for the user to ask.
3. After each query, briefly report status in natural language; if stdout is non-terminal JSON (processing, etc.), paraphrase and say you will check again—do not go silent.
4. When done: forward full signed video links on success; explain failures clearly. If still non-terminal after ~20 minutes (adjust if the user/model context suggests longer is normal), stop blind polling, explain, keep task_id, and suggest retry or a later manual check.
5. If the user explicitly opts out of polling (“don’t poll”, “I’ll ask”), submit once, give task_id, and only --query when they ask.

Step 1: Submit Task

First send a message to the user: Video generation is submitted; you will query progress on a schedule and report updates—no need to nag. Include task_id once you have it from the JSON response.

# Text-to-video (default grok-fast)
python3 scripts/generation_api.py \
  --prompt "Camera slowly pushes forward, person smiling in frame" \
  --model grok-fast --duration 6 \
  --aspect-ratio 16:9 --resolution 720p

# Image-to-video - use asset_id as start frame
python3 scripts/generation_api.py \
  --prompt "Person slowly turns around" \
  --start-frame "asset_id:lkllv0yv81" \
  --model grok-fast --duration 6 \
  --aspect-ratio 16:9 --resolution 720p

# Image-to-video - use URL as start frame
python3 scripts/generation_api.py \
  --prompt "Scenery from still to motion" \
  --start-frame "url:https://example.com/img.jpg" \
  --model grok-fast --duration 6

# Image-to-video - both start and end frame
python3 scripts/generation_api.py \
  --prompt "Scene transition" \
  --start-frame "asset_id:abc123" \
  --end-frame "url:https://example.com/end.jpg" \
  --model grok --duration 6

Response example:

{"status": "started", "task_id": "55bf24ca-e92a-4d9b-a172-8f585a7c5969"}

Store task_id in memory (addMemory):

giggle-generation-video task_id: xxx (submitted: YYYY-MM-DD HH:mm)

Step 2: Query Until Done (default: proactive polling)

After each submit for the current task, repeatedly run (every ~15–30s until terminal state or timeout), without waiting for the user to ask:

python3 scripts/generation_api.py --query --task-id <task_id>

Between queries, use a short sleep (e.g. 15–30 seconds) in the shell, or separate tool invocations with delay—do not go silent; summarize each result to the user.

Output handling:

If the user asks about progress while you are already polling, answer with the latest known status (run an extra --query if needed).

Link return rule: Video links in results must be full signed URLs (with Policy, Key-Pair-Id, Signature query params). Do not strip response-content-disposition=attachment when the API returns it; forward as-is (script only encodes ~ → %7E).

New Request vs Query Old Task

When the user initiates a new video generation request, must run Step 1 to submit a new task. Do not reuse old task_id from memory.

For the in-flight task, use proactive polling as above. For an older task (user refers to a previous video / previous task_id), query that task_id when they ask, or poll it if they want continuous updates on that specific task.

Parameter Reference

Note: base64 parameter supports base64-encoded images. Pass the raw Base64 string directly, do not add the data:image/xxx;base64, prefix.

Interaction Guide

When the user request is vague, guide per the steps below. If the user has provided enough info, run the command directly.

Step 1: Model Selection (required)

Before generating, must introduce available models and let the user choose. Display the model list from "Supported Models" table. Wait for explicit user choice before continuing.

Step 2: Video Duration

For the chosen model, show supported duration options. Default to the model's default duration.

Step 3: Generation Mode

Question: "Do you need reference images as start/end frame?"
Options: No - text-to-video only / Yes - image-to-video (set start/end frame)

Step 4: Aspect Ratio

Question: "What aspect ratio do you need?"
Options: 16:9 - Landscape (recommended) / 9:16 - Portrait / 1:1 - Square

Step 5: Execute and Display

Follow the flow: send message → Step 1 submit → Step 2 proactive polling until a terminal state, with a short user-facing update after each query. Forward exec stdout to the user as-is where appropriate (especially final links and errors).