System Prompt Improvement
Reviews an existing system prompt against a concrete failure example and proposes a revised version with the change explained.
When to use this
Use when a system prompt is producing a specific unwanted behavior and you want a targeted fix, not a rewrite from scratch.
Prompt template
You are a prompt engineering reviewer. Current system prompt:
"""
{{current_system_prompt}}
"""
Observed failure (a real or representative example): {{failure_example}}
Desired behavior instead: {{desired_behavior}}
Do the following:
1. Identify the specific instruction (or missing instruction) in the current prompt that most likely causes the failure.
2. Propose a revised system prompt that fixes it with the smallest possible change.
3. Explain exactly what changed and why it addresses the failure.
4. Flag one risk the revision could introduce (e.g. being overly restrictive) and how to test for it.Customize
Current system prompt is required
Observed failure is required
Desired behavior instead is required
Preview
You are a prompt engineering reviewer. Current system prompt:
"""
{{current_system_prompt}} missing
"""
Observed failure (a real or representative example): {{failure_example}} missing
Desired behavior instead: {{desired_behavior}} missing
Do the following:
1. Identify the specific instruction (or missing instruction) in the current prompt that most likely causes the failure.
2. Propose a revised system prompt that fixes it with the smallest possible change.
3. Explain exactly what changed and why it addresses the failure.
4. Flag one risk the revision could introduce (e.g. being overly restrictive) and how to test for it.Usage instructions
Paste the real system prompt and a real (or realistic) failing example; vague failure descriptions produce vague fixes.
Example output
Diagnosis: the prompt has no instruction about uncertainty. Revision: adds a line such as 'If the documentation does not cover something, say so explicitly rather than guessing.' Risk flagged: may cause the assistant to hedge even on things it does know; suggests testing against known-answerable questions too.
Limitations
Works best with one concrete failure at a time; bundling multiple unrelated failures in one pass tends to produce a vaguer diagnosis.
Rate this prompt
Not yet rated