inference-sh skills

For versioned software, use the MAJOR.MINOR.PATCH format:

For SaaS products with continuous deployment, Date-Based Versioning (e.g., 2026-02-08) is often preferred.

An anti-persona represents the segment of people who are NOT your customers. Defining an anti-persona helps prevent wasted product effort on customers you cannot serve (e.g., an enterprise customer when you are a self-serve SMB SaaS).

Example Anti-Persona: ANTI-PERSONA: "Enterprise Earl"

• Profile: CTO at a 5,000+ person enterprise.
• Needs: SOC 2, HIPAA, on-premise deployment, 18-month procurement cycles.
• Why NOT: Product is self-serve SaaS for SMB/mid-market; enterprise needs require too much investment.

A successful Twitter thread follows a specific anatomy to maximize engagement and readability. Use this template for structuring your content:

1. Tweet 1 (Hook): A bold claim followed by a thread emoji (e.g., 🧵). This must work as a standalone tweet.
2. Tweet 2: Provide context or explain why the topic matters.
3. Tweets 3-9 (Content): One specific point per tweet. Use numbering (e.g., 1/, 2/) to signal progress.
4. Tweet 10: A summary or the biggest takeaway.
5. Tweet 11 (CTA): A Call to Action (e.g., follow, retweet, or bookmark).

The SDK supports several ways to provide file data in the input dictionary of a run() call:

1. File Path: A string representing the local path.
2. Data URI (Base64): A string in the format data:<mime_type>;base64,<data>.
3. Remote URLs: A direct URL (e.g., https://...) which is used without requiring an upload.
4. Bytes/File Objects: When using upload_file(), you can pass raw bytes or a file-like object.

# Data URI (Base64)
import base64
with open("image.png", "rb") as f:
    b64 = base64.b64encode(f.read()).decode()

result = client.run({
    "app": "processor",
    "input": {"image": f"data:image/png;base64,{b64}"}
})

# Remote URL
result = client.run({
    "app": "image-processor",
    "input": {"image": "https://example.com/image.png"}
})

Most products should target 2-4 personas. Attempting to serve more than 4 often leads to diluted focus.

The prompt_influence parameter (0.0 to 1.0) determines how strictly the model follows your text description. Use lower values for creative/surprising results and higher values for precise/literal sound reproduction.

# Loose interpretation - creative result
belt app run elevenlabs/sound-effects --input '{
  "text": "Magical fairy dust sparkle",
  "prompt_influence": 0.1
}'

# Literal interpretation - precise result
belt app run elevenlabs/sound-effects --input '{
  "text": "Single gunshot, pistol, indoor range",
  "prompt_influence": 0.8
}'

Follow these rules to avoid video distortion and artifacts:

1. The Golden Rule: Subtle > Dramatic. Requesting high-intensity movement often causes warping. Use "slowly walking" instead of "running and jumping".
2. Prompt Structure: Use the pattern [Camera movement] + [Subject motion] + [Atmospheric effects] + [Mood/pace].
3. Duration: Aim for 2-5 seconds for the highest quality. Avoid clips longer than 10 seconds as quality degrades significantly.

When building automated workflows, follow these best practices:

1. Rate limiting: Add delays between API calls to avoid being throttled.
2. Error handling: Always check return codes of your commands.
3. Logging: Track all operations for debugging and auditing.
4. Idempotency: Design workflows so they can be safely re-run without side effects.
5. Monitoring: Implement alerts for failures.
6. Backups: Save intermediate results (e.g., to JSON files).
7. Timeouts: Set reasonable limits on how long a task should run.