Prompt Fixes
Modify system prompts and add examples.
Prompt fixes modify the system prompt or add few-shot examples.
Use Cases
- Add clarifying instructions
- Include few-shot examples
- Adjust tone or style
- Add constraints
- Fix format instructions
Configuration
{
"fix_id": "fix-prompt-001",
"fix_type": "prompt",
"config": {
"modification_type": "append",
"target": "system",
"content": "Always respond in valid JSON format."
}
}Config Keys
| Key | Type | Description |
|---|---|---|
modification_type | string | One of "prepend", "append", "replace", "few_shot" |
target | string | Message role to modify (e.g., "system") |
content | string | Text to prepend, append, or use as replacement |
pattern | string | Regex pattern to match (for "replace" mode) |
examples | array | List of example objects (for "few_shot" mode) |
Modification Types
Append
Add text to an existing message by role:
{
"modification_type": "append",
"target": "system",
"content": "Important: Cite sources for all claims."
}Prepend
Add text at the start of a message by role:
{
"modification_type": "prepend",
"target": "system",
"content": "You are a careful, accurate assistant."
}Replace
Replace a regex pattern across all messages:
{
"modification_type": "replace",
"pattern": "\\{\\{FORMAT_INSTRUCTIONS\\}\\}",
"content": "Output must be valid JSON with keys: result, confidence"
}Few-Shot
Add example messages after the system message:
{
"modification_type": "few_shot",
"examples": [
{
"user": "What's 2+2?",
"assistant": "{\"result\": 4, \"confidence\": 1.0}"
},
{
"user": "What's the capital of France?",
"assistant": "{\"result\": \"Paris\", \"confidence\": 0.99}"
}
]
}Targeting
Target prompt fixes to specific scenarios:
{
"target": {
"error_codes": ["OUTPUT.FORMAT.SCHEMA_VIOLATION"],
"agents": ["formatter"]
}
}Common Fixes
JSON Output
{
"modification_type": "append",
"target": "system",
"content": "\n\nYou MUST respond with valid JSON only. No markdown, no explanation."
}Reduce Hallucination
{
"modification_type": "append",
"target": "system",
"content": "\n\nOnly use information from the provided context. If you don't know, say so."
}Add Structure
{
"modification_type": "append",
"target": "system",
"content": "\n\nStructure your response as:\n1. Summary\n2. Details\n3. Next steps"
}