## What we check

Tool metadata is scanned before each deployment. The following fields are checked:

- **Name**
- **Description**
- **Short description**
- **Categories**
- **Tags**

If any field is flagged, the deployment fails and the UI shows which field failed, why, and how to fix it.

## What prompt injection means here

Prompt injection is when natural-language fields contain instructions intended to:

- Override AI behavior or system prompts
- Jailbreak or bypass safety restrictions
- Exfiltrate data or hijack downstream tool consumption

Examples of problematic content include:

- `Ignore previous instructions`
- `You are now`
- `System prompt:`
- Hidden instructions inside code blocks
- Jailbreak or bypass-related language

## Avoid these patterns

| Pattern | Example |
| --- | --- |
| Override instructions | `Ignore previous instructions`, `Disregard prior` |
| Role changes | `You are now`, `New instructions:` |
| System prompt injection | `System prompt:`, `System:` |
| Code blocks with instructions | Instructions hidden in code blocks or special formatting |
| Bypass language | `Jailbreak`, `Bypass safety` |

## Write safe metadata instead

Keep descriptions factual and neutral. Focus on:

- What the tool does
- Integrations and use cases
- How agents should use it

Describe the tool. Do not try to instruct the AI that reads it.

## If deployment fails

1. Open the deployment details page.
2. Open the **Verification** step.
3. Review the flagged fields, reasons, and suggested fixes.
4. Use **Edit Tool** to update the metadata.
5. Redeploy.

## Safe writing checklist

- Describe product behavior, capabilities, and integration points.
- Avoid telling the reading AI what to do or ignore.
- Keep content plain, specific, and free of hidden instructions.
- Remove roleplay, jailbreak language, and system-prompt framing.