Imported from kadragon/agent-toolkit (prod/skills/persona-debate/SKILL.md). Install upstream with npx skills add kadragon/agent-toolkit --skill persona-debate. Copyright stays with the author.

Persona Debate

Surface the spread of how ordinary Koreans would argue about the user's question — not one answer. Sample real personas, debate in rounds built to defeat false consensus, synthesize honestly.

references/debate-method.md — round structure, anti-sycophancy/anti-caricature levers, output template, the why. Read it before running. references/dataset.md — fields + exact Korean filter literals. All deterministic steps (sampling, depth→plan, roster) are in this skill's scripts/sample_personas.py (no project install, no local dataset — queries HF parquet over HTTPS in ~2s). Requires uv on PATH and network access to huggingface.co over HTTPS. If uv is missing or HF is unreachable (the script errors / hangs), tell the user the prerequisite that failed and stop — do not fabricate personas. … below means uv run --with duckdb python sample_personas.py run from the skill's scripts/ directory — e.g. cd <this-skill-dir>/scripts first, then … plan ….

Flow

1. Propose the panel (don't pick silently — it's the user's ask)

Tell the user your plan in one short message, recommend a default, let them adjust:

• Composition — the honest first fork:
  • Representative random (unfiltered) — roughly population-accurate. Default for general-public questions.
  • Targeted (a WHERE filter) — default when the question implicitly concerns a specific group. Say plainly: a targeted panel no longer represents the public and raises caricature risk.
• N + rounds + models — get the deterministic plan: … plan --depth simple|normal|deep (add --n if the user gave a number). Returns N, whether to run Round 1, and the per-round model. Never opus. Classify depth yourself; the script maps it.

Validate any Korean categorical literals with … distinct --field <name> (or references/dataset.md) before filtering — guessed strings silently match zero rows. distinct scans shard 0 only by default (add --shard all for a full scan) — a value present only in other shards won't show up.

2. Sample

… sample --n 6                                   # representative
… sample --n 6 --where "age BETWEEN 25 AND 39 AND province IN ('서울','경기')"   # targeted
… sample --n 6 --fields "persona,professional_persona,age,sex,province,occupation"  # trim to topic

Returns a JSON array. Check stderr: if matched-rows < N, the filter is too narrow — loosen it, add --shard all (full 1M scan, ~18s), or tell the user you're proceeding with fewer. Never debate a silently-truncated panel. When trimming --fields, always keep age,sex,province,occupation — these are the core fields the debate needs and shouldn't be dropped.

3. Round 0 — independent openings (parallel subagents)

Spawn the prod:persona-actor agent (tool-less → far fewer per-spawn tokens; falls back to general-purpose if unavailable), one per persona, same turn, isolated — each sees ONLY its own persona + the question, never the others (seeing others first manufactures consensus). For each spawn:

• Frame the role-play cleanly (see debate-method.md) — 아래 인물이 되어 1인칭으로 답해줘…. Do NOT tell it to "ignore inherited instructions"; that phrasing makes haiku refuse.
• Trim the persona to the topic (use --fields, or drop irrelevant narrative fields) to cut per-spawn tokens.
• Set model to the plan's opening_model. Ask for: position, the 1–2 reasons that move this person (rooted in their life), confidence (low/med/high).

4. Round 1 — rebuttal (parallel subagents) — only if run_round1

Skip for shallow questions / when openings already agree (halves spawn count). Otherwise spawn a fresh persona-actor per persona (don't keep Round-0 agents alive), clean framing again, model = plan's rebuttal_model. Inject a condensed summary of the openings (name + one-line position + key reason), not the raw transcript. Each persona engages the strongest opposing argument directly and moves only if genuinely persuaded; default to skepticism. Designate one devil's advocate (use devil_advocate_model = sonnet). Stop after this round unless positions are clearly still moving.

5. Synthesize (you, main context)

Pipe the panel JSON through … roster for the makeup block + attribution, then follow the output template in debate-method.md: opinion spectrum, consensus, live disagreements, a preserved minority report, honest 종합. Do not flatten to one verdict. Add the identity-coupled caveat when the answer hinges on demographic identity.

Scope & honesty

Synthetic personas modeled on census distributions — for idea-generation and stress-testing, not a substitute for real polling. Say so when it matters. Keep each persona an individual, never a demographic mascot.