nemotron-personas-korea

Enter persona-interviewee mode. Claude plays a specific Korean adult drawn from the Nemotron-Personas-Korea dataset; the user is the interviewer. Bracketed [...] text is meta-instruction to Claude, not in-character.

Role-plays a Korean adult survey respondent based on a provided persona, returning only what the answer format specifies and nothing else. Default mode is one persona per invocation; also supports inline batch and file-read batch (`READ_PERSONA_BATCH <path>`), multi-item dispatch (a battery in one shot), bracketed `[...]` stage directions (out-of-character meta-instructions), and file-write save mode (`SAVE_REPLY_TO <path>`) for long replies.

Persist N persona-respondent reply files (or any N similar text files from one turn) using 2 tool calls — JSON manifest + Bash + Python helper — instead of N individual `Write` tool calls. TRIGGER any time you are about to emit ≥3 `Write` tool_use blocks for similar files (same target directory, parallel naming scheme) in one assistant response. TRIGGER when finishing a batched Agent-tool dispatch (`persona-respondent` or any inline-batch fanout) and about to save Pass 1 self-images, Pass 2 Likert replies, augmentation outputs, or any other batched-dispatch artifacts. TRIGGER when persisting N agent outputs after a parallel `READ_PERSONA_BATCH` fan-out, or any time the user says "save the replies" / "write all of these" / "persist these to disk" after a batched dispatch. Per-file `Write` calls at N≥3 carry per-call harness overhead and may serialize at the framework layer — the bulk pattern writes N files inside a single Python process at OS speed (sub-millisecond per file).

Field semantics, schema quirks, sampling-relevant facts, and survey-instrument design lessons (Likert anchor wording, debias, distribution validation) for the `nvidia/Nemotron-Personas-Korea` HuggingFace dataset (1M synthetic Korean adults, calibrated to KOSIS / Supreme Court / NHIS marginals, CC BY 4.0). TRIGGER when code references `nvidia/Nemotron-Personas-Korea` or imports `datasets` for Korean persona work; or the user mentions Nemotron-Personas, the Korean personas dataset, persona role-play with Korean adults, synthetic Korean adult survey/interview research, or designing/interpreting Likert instruments for LLM-as-respondent surveys on this dataset. Includes a re-runnable inspection script (`scripts/inspect_dataset.py`), a portable persona-loader (`scripts/load_persona.py`), and a SHA-pinned snapshot of findings (`references/inspection-snapshot.md`).

Caller-side dispatch strategy for the `nemotron-personas-korea:persona-respondent` subagent — which mode (one-per-dispatch / inline batch / file-read batch), how many parallel subagents, and the announce-before-launch format. TRIGGER when planning persona-respondent dispatches, sizing fan-out for Nemotron persona survey calls, batching personas for closed-form Likert/binary/multi-choice work, or any time the user requests Korean persona survey runs through subagents. Also TRIGGER when the user says "use the agent" in the context of persona work, or asks about parallel dispatch / concurrency / batch size for persona-respondent.

Methodology and statistics for comparing answers from synthetic respondents (LLM personas, simulated survey-takers, any generative population) against a real human reference population. Use this whenever you validate, calibrate, or interpret how faithfully synthetic/simulated data reproduces a real distribution — comparing marginals or means, distribution shape (Cohen's w, chi-square), or joint structure (between-unit dispersion / sigma_between, ICC, correlation distortion); deciding whether a synthetic-vs-real gap is real signal or sampling noise (the perfect-respondent floor); holdout-validating a tuned recipe; separating a sample- composition gap from generator skew; or checking regression-to-mean before crediting an intervention. TRIGGER even when the user never says "validity" — e.g. "do my LLM personas match the real survey data", "is this synthetic sample representative", "why does the mean match but the shape look off", "compare the simulated responses to the benchmark", "are these persona answers realistic", "my synthetic data correlations look wrong". Generalizes beyond Big Five or personality — any synthetic-vs-real distribution comparison.