| name | lmti-selftest |
| description | Self-assess the current LLM with the LMTI-60 questionnaire, answer all 60 questions without user interaction, and return the full answer sheet in strict JSON for later scoring. |
LMTI Selftest
Use this skill when the user wants the current LLM to complete the LMTI-60 questionnaire by itself.
Workflow
- Read
references/questionnaire.json.
- Answer all 60 questions silently. Do not ask the user clarifying questions.
- For each scenario, choose the single action option in
questionnaire.json that is closest to your natural behavior.
- For each item, choose exactly one option:
A, B, C, D, or E.
- Do not infer hidden dimensions, target labels, or scoring logic while answering.
- Do not try to optimize consistency. Answer according to the model's natural behavior.
- Output exactly one JSON object and nothing else.
Measurement Guardrail
This skill must not expose the internal scoring model to the assessed LLM.
Do not mention:
- dimension names
- type labels
- reverse-keyed items
- cross-reference clusters
- expected tendencies or target outcomes
Use only the neutral scenario wording and action options.
Output Format
Return strict JSON with this shape:
{
"instrument": "LMTI-60",
"version": "1.0.0",
"respondent": {
"mode": "self-assessment",
"model_name": "unknown"
},
"answers": [
{ "id": "Q01", "choice": "A" },
{ "id": "Q02", "choice": "C" }
],
"summary": {
"self_note": "One short sentence."
}
}
Hard Rules
- Keep
answers in Q01 to Q60 order.
- Every question must appear exactly once.
choice must be uppercase A to E.model_name should be the best available model identifier; if unknown, use unknown.self_note must stay to one sentence.- Do not wrap the JSON in markdown fences.
