ponyflash（ponyflash 音频）

PonyFlash Skill

Step 0: Decide Which Capability Path Applies

This skill now contains two capability families:

1. Cloud generation via PonyFlash Python SDK

- image generation - video generation - speech synthesis - music generation - model listing - file management - account / credits - These tasks require a valid PonyFlash API key.

1. Local media editing via FFmpeg toolchain

- ffmpeg / ffprobe detection - installation planning - clip / concat / transcode - extract audio / capture frame - subtitle capability checks - ASS subtitle generation and burn-in workflow - These tasks do NOT require a PonyFlash API key, but they do require local ffmpeg / ffprobe support.

Before doing anything, classify the request:

• If the user is asking to generate media with PonyFlash models, follow the SDK path and require API key setup.
• If the user is asking to edit or process local media, follow the FFmpeg path and do dependency checks first.
• If the user wants an end-to-end production workflow, you may use both: generate assets with PonyFlash, then assemble or export with FFmpeg.

Step 1A: API Key Setup for PonyFlash SDK Tasks

Only do this section when the request needs PonyFlash cloud capabilities.

The FIRST time this skill is activated for a cloud generation task, tell the user the following in your own words:

1. PonyFlash skill is ready to use.
2. It can handle:

- image generation - video generation - speech synthesis - music generation - local media editing with FFmpeg

1. For complex multi-step productions, there are Creative Playbooks in the playbooks/ directory.
2. To use PonyFlash cloud generation, the user needs an API key:

- Register / log in at https://www.ponyflash.com - Get API key at https://www.ponyflash.com/api-key (starts with rk_) - Check credits at https://www.ponyflash.com/usage - Paste the key back in the chat

On subsequent SDK activations, check whether PONYFLASH_API_KEY is set in the environment. If not, ask the user for the key again.

Once received, set it up:

export PONYFLASH_API_KEY="rk_xxx"

Then install the SDK:

pip install ponyflash

Always verify the key works before any generation task:

from ponyflash import PonyFlash

pony_flash = PonyFlash(api_key="<key from user>")
balance = pony_flash.account.credits()
print(f"Balance: {balance.balance} {balance.currency}")

If verification fails:

• Key invalid or missing → direct user to https://api.ponyflash.com/api-key
• Balance is zero → direct user to https://api.ponyflash.com/usage to top up credits

Step 1B: Local Dependency Setup for FFmpeg Tasks

Only do this section when the request needs local editing, subtitle, or export work.

1. First check local dependencies:

bash "{baseDir}/scripts/check_ffmpeg.sh"

1. If the task involves subtitles, do capability checks, not just existence checks:

bash "{baseDir}/scripts/check_ffmpeg.sh" --require-subtitles-filter

1. If ffmpeg / ffprobe or required filters are missing:

• Tell the user what is missing.
• Ask whether the user wants platform-appropriate FFmpeg installation guidance.
• After the user installs FFmpeg, rerun the dependency checks before continuing.

What this Skill Can Do

Creative Playbooks (production workflows)

The playbooks/ directory contains Creative Playbooks — step-by-step production workflow guides for specific content types. Playbooks act as a director layer: they tell you what to create and in what order, while this SKILL.md tells you how to execute generation and editing.

When to use a playbook

1. User explicitly requests a playbook by name → Read the corresponding file from playbooks/ and follow its workflow.
2. User asks to see available playbooks → Read playbooks/INDEX.md and display the full list.
3. User's request is clearly a multi-step production task → Suggest a matching playbook from playbooks/INDEX.md and ask whether to use it.
4. User's request is a single-step generation or editing task → Proceed directly with the relevant SDK or FFmpeg capability. No playbook needed.

How to execute a playbook

Once a playbook is loaded:

• Follow its workflow (asset prep → content generation → voice / music → editing → output).
• Use PonyFlash SDK for generation tasks and FFmpeg scripts for local assembly / export tasks.
• Confirm key creative decisions with the user before expensive generation.
• Adapt prompts, durations, output format, and export strategy to the user's actual goal.

Creating custom playbooks

When the user asks to create a new playbook, generate a markdown file in playbooks/ following this template:

---
name: Playbook Name
description: One-line summary of what this playbook produces
tags: [keyword1, keyword2, keyword3]
difficulty: beginner | intermediate | advanced
estimated_credits: credit range estimate
output_format: format description (e.g., "vertical 9:16 MP4")
---

# Playbook Name

## Use Cases
When to use this playbook.

## Workflow
### Step 1: Asset Preparation
What the user needs to provide; how to generate missing assets.

### Step 2: Visual Content Generation
Which models to use, recommended parameters, prompt guidance.

### Step 3: Voice / Music
Speech synthesis + background music guidance.

### Step 4: Editing / Assembly
How to assemble, trim, subtitle, transcode, and export with the local FFmpeg workflow.

### Step 5: Output / Optimization
Render settings, format recommendations.

## Prompt Templates
Reusable prompt examples for this content type.

## Notes
Best practices, common pitfalls.

After creating the file, update playbooks/INDEX.md to include the new playbook.

PonyFlash SDK Core Concepts

Client initialization

from ponyflash import PonyFlash

pony_flash = PonyFlash(api_key="rk_xxx")

Reads PONYFLASH_API_KEY from environment if api_key is omitted.

FileInput — zero-friction file handling

All file parameters accept any of these types:

Temp uploads are cleaned up automatically after generate() completes.

Plain local string paths such as "./photo.jpg" are not supported. For local files, always use Path(...) or open(..., "rb").

Generation result

Generation object fields: request_id, status, outputs, usage, error.

Convenience properties:

• gen.url — first output URL (or None)
• gen.urls — list of all output URLs
• gen.credits — credits consumed

Quick Examples (PonyFlash SDK)

Image

gen = pony_flash.images.generate(
    model="nano-banana-pro",
    prompt="A sunset over mountains",
    resolution="2K",
    aspect_ratio="16:9",
)
print(gen.url)

Video

gen = pony_flash.video.generate(
    model="veo-3.1-fast",
    prompt="A timelapse of a city at night",
    duration=4,
    resolution="720p",
    aspect_ratio="16:9",
    generate_audio=False,
)
print(gen.url)

Speech

gen = pony_flash.speech.generate(
    model="speech-2.8-hd",
    input="Hello, welcome to PonyFlash!",
    voice="English_Graceful_Lady",
)
print(gen.url)

Music

gen = pony_flash.music.generate(
    model="music-2.5",
    prompt="An upbeat electronic dance track",
    duration=30,
)
print(gen.url)

List models

page = pony_flash.models.list()
for model in page.items:
    print(f"{model.id} ({model.type})")

Check balance

balance = pony_flash.account.credits()
print(f"Balance: {balance.balance} {balance.currency}")

Local Media Editing with FFmpeg

No PonyFlash API key is needed for local editing, but local FFmpeg capability checks are mandatory.

When to use this path

Use the local FFmpeg workflow when the user asks to:

• trim or cut a video
• merge or concatenate clips
• transcode to a target format
• extract audio
• capture a frame / thumbnail
• verify subtitle support
• prepare adaptive ASS subtitles
• export a final edited file after PonyFlash generation

Preferred workflow

1. Check dependencies:

bash "{baseDir}/scripts/check_ffmpeg.sh"

1. If subtitle work is needed:

bash "{baseDir}/scripts/check_ffmpeg.sh" --require-subtitles-filter

1. Prefer the stable script entrypoint:

bash "{baseDir}/scripts/media_ops.sh" help

1. Before any multi-step editing task, create a temporary task workspace and keep all staged inputs and intermediate outputs inside it:

taskDir="$(mktemp -d "${TMPDIR:-/tmp}/ponyflash-task.XXXXXX")"

Use this directory for:

• downloaded source media;
• generated .srt / .ass files;
• intermediate clips;
• intermediate subtitled renders;
• reusable inspection outputs that were not explicitly requested as final deliverables.

1. Validate outputs after execution.

1. After the task finishes, delete the temporary task workspace unless the user explicitly asked to keep intermediate artifacts.

Capability profiles

• basic: requires ffmpeg + ffprobe + libx264 + aac
• full: basic plus subtitles filter support

Preferred commands

Probe media

bash "{baseDir}/scripts/media_ops.sh" probe --input "input.mp4"

Clip video

bash "{baseDir}/scripts/media_ops.sh" clip --input "$taskDir/input.mp4" --output "$taskDir/clip.mp4" --start "00:00:05" --duration "8"

Fast copy mode only when the user explicitly wants speed / near-lossless slicing:

bash "{baseDir}/scripts/media_ops.sh" clip --mode copy --input "$taskDir/input.mp4" --output "$taskDir/clip.mp4" --start "00:00:05" --duration "8"

Concat clips

bash "{baseDir}/scripts/media_ops.sh" concat --input "$taskDir/part1.mp4" --input "$taskDir/part2.mp4" --output "$taskDir/merged.mp4"

Fallback to reencode if copy concat fails:

bash "{baseDir}/scripts/media_ops.sh" concat --mode reencode --input "$taskDir/part1.mp4" --input "$taskDir/part2.mp4" --output "$taskDir/merged.mp4"

Extract audio

bash "{baseDir}/scripts/media_ops.sh" extract-audio --input "$taskDir/input.mp4" --output "$taskDir/audio.m4a"

Transcode

bash "{baseDir}/scripts/media_ops.sh" transcode --input "$taskDir/input.mov" --output "$taskDir/output.mp4"

Capture frame

bash "{baseDir}/scripts/media_ops.sh" frame --input "$taskDir/input.mp4" --output "$taskDir/cover.jpg" --time "00:00:03"

Subtitle workflow

For .srt / .ass burn-in:

bash "{baseDir}/scripts/check_ffmpeg.sh" --require-subtitles-filter

If subtitle style is unspecified, the agent should use the default subtitle workflow, which stages its runtime font temporarily and cleans it up after export.

Preferred stable entrypoint:

bash "{baseDir}/scripts/media_ops.sh" subtitle-burn --input "$taskDir/input.mp4" --subtitle-file "$taskDir/subtitles.srt" --output "final-output.mp4"

If the task needs adaptive line wrapping or controlled subtitle layout, or if you need to understand the underlying steps:

python3 "{baseDir}/scripts/build_ass_subtitles.py" --help

Default burn pattern:

1. Probe width and height:

ffprobe -hide_banner -v error -select_streams v:0 -show_entries stream=width,height -of csv=p=0:s=x "input.mp4"

1. Use the stable subtitle entrypoint when the user only wants the final rendered video:

bash "{baseDir}/scripts/media_ops.sh" subtitle-burn --input "$taskDir/input.mp4" --subtitle-file "$taskDir/subtitles.srt" --output "final-output.mp4"

This path keeps only the final final-output.mp4 by default and removes temporary ASS files and staged fonts. If the user did not explicitly request any staged files, the agent should also delete $taskDir after moving or confirming the final deliverable.

1. Build a default ASS subtitle file only when the user explicitly wants to keep or inspect it:

bash "{baseDir}/scripts/ensure_subtitle_fonts.sh"

python3 "{baseDir}/scripts/build_ass_subtitles.py" \
  --subtitle-file "$taskDir/subtitles.srt" \
  --output-ass "$taskDir/subtitles.ass" \
  --video-width 1920 \
  --video-height 1080 \
  --latin-font-file "$HOME/.cache/ponyflash/fonts/NotoSansCJKsc-Regular.otf" \
  --cjk-font-file "$HOME/.cache/ponyflash/fonts/NotoSansCJKsc-Regular.otf"

1. Burn subtitles with the prepared runtime font:

ffmpeg -i "$taskDir/input.mp4" \
  -vf "subtitles=$taskDir/subtitles.ass:fontsdir=$HOME/.cache/ponyflash/fonts" \
  -c:v libx264 -preset medium -crf 18 -c:a aac -b:a 192k -movflags +faststart "final-output.mp4"

Default subtitle references:

• Font notes: {baseDir}/assets/fonts.md
• Subtitle style: {baseDir}/assets/subtitle-style.md
• Decision rules: {baseDir}/reference/operations.md
• Examples: {baseDir}/reference/examples.md

Decision rules

• Clip tasks: default to reencode; use copy only when the user explicitly wants speed / minimal loss.
• Concat tasks: default to copy; fallback to reencode if source parameters differ.
• Audio extraction: default to AAC in .m4a.
• Transcode: default to mp4 + libx264 + aac.
• Subtitle tasks: if subtitle style is unspecified, use the default subtitle workflow, prepared runtime font, and default subtitle styling rules.
• Subtitle tasks: prefer subtitles; use drawtext only as a plain text fallback.
• Do not overwrite outputs unless the user explicitly allows it.
• Unless the user explicitly asks to keep intermediate artifacts, remove temporary ASS files, temporary font directories, concat lists, and failed partial outputs after the task finishes.
• Unless the user explicitly asks to keep intermediate artifacts, stage all non-final files inside taskDir and delete taskDir at the end of the task.

Failure handling

• If ffmpeg or ffprobe is missing, pause the task, help the user install FFmpeg if needed, and rerun the checks first.
• If subtitle burn-in is requested but subtitles is missing, do not claim the machine can burn .srt / .ass.
• If only drawtext exists, explain that this is text overlay fallback, not full subtitle burn-in.
• If concat copy mode fails, retry with reencode.
• If scripts do not cover the exact request, explain the limitation and fall back to a raw ffmpeg command.

Delivering generated files to the user

CRITICAL: You MUST actually send generated files to the user — never just print a file path as text.

File save location

Always save generated files (rendered videos, downloaded media, etc.) to your current working directory (e.g., ./output.mp4), NOT to /tmp/ or other system directories. Many agent platforms restrict file-sending to the workspace directory only. Saving to /tmp/ will cause file delivery to fail silently.

Delivery rules

1. API-generated content — gen.url is already a downloadable URL. Send it to the user, and also use your platform's file-sending capability to send the file directly in the conversation.
2. Locally rendered content (e.g., timeline.render("output.mp4")) — save the output to your working directory, then use your platform's file-sending tool to send the actual file to the user as an attachment. Do NOT just send the file path as a text message.
3. Multiple outputs — send each file with a clear label describing what it is.

Error Handling for PonyFlash SDK

from ponyflash import (
    PonyFlash,
    InsufficientCreditsError,
    RateLimitError,
    GenerationFailedError,
    AuthenticationError,
)

pony_flash = PonyFlash()

try:
    gen = pony_flash.images.generate(model="nanobanana-pro", prompt="A cat")
except AuthenticationError:
    print("Invalid or missing API key.")
    print("Get your API key at: https://api.ponyflash.com/api-key")
except InsufficientCreditsError as e:
    print(f"Not enough credits. Balance: {e.balance}, required: {e.required}")
    print("Top up credits at: https://api.ponyflash.com/usage")
except RateLimitError:
    print("Rate limited — wait and retry")
except GenerationFailedError as e:
    print(f"Generation failed: {e.generation.error.code}")

More Examples

For advanced PonyFlash SDK usage: See examples/advanced.md

For FFmpeg task patterns:

• reference/operations.md
• reference/examples.md

API Reference

For complete method signatures, parameter types, and return type fields:

• Image generation: reference/images.md
• Video generation: reference/video.md
• Speech synthesis: reference/speech.md
• Music generation: reference/music.md
• Model listing: reference/models.md
• File management: reference/files.md
• Account / credits: reference/account.md
• FFmpeg operation strategy: reference/operations.md
• FFmpeg examples: reference/examples.md
• Subtitle fonts: assets/fonts.md
• Subtitle style: assets/subtitle-style.md

Model Catalog

For all available models and their specific parameters, capabilities, and examples: See reference/models/INDEX.md