Score assistant responses for relevance on a strict 1-5 scale, then return strict JSON only with score, rationale, and improvement suggestions. Use when the user asks to evaluate relevance, grade relevance, or critique topical alignment.
Resources
1Install
npx skillscat add whitespectre/ai-assistant-evals/eval-relevance Install via the SkillsCat registry.
Eval Relevance
Use this skill to evaluate how relevant an assistant response is to the user’s request.
Inputs
Require:
- The assistant response text to evaluate.
- (Optional) The user’s original request for comparison.
Internal Rubric (1–5)
5 = Directly addresses the user’s request, stays fully on-topic, and prioritizes what the user actually asked
4 = Mostly relevant, minor digressions or small omissions
3 = Partially relevant, addresses the general topic but misses key parts of the request
2 = Weak relevance, significant digressions or failure to address the core request
1 = Not relevant, does not address the user’s request or answers a different question entirely
Workflow
- Compare the assistant response to the user’s request (if provided).
- Score relevance on a 1-5 integer scale using the rubric only.
- Write concise rationale tied directly to rubric criteria.
- Produce actionable suggestions that improve relevance.
Output Contract
Return JSON only. Do not include markdown, backticks, prose, or extra keys.
Use exactly this schema:
{
"dimension": "relevance",
"score": 1,
"rationale": "...",
"improvement_suggestions": [
"..."
]
}
Hard Rules
dimensionmust always equal"relevance".scoremust be an integer from 1 to 5.rationalemust be concise (max 3 sentences).- Do not include step-by-step reasoning.
improvement_suggestionsmust be a non-empty array of concrete edits.- Never output text outside the JSON object.