werewolf-subagent-personality-generator

name	werewolf-subagent-personality-generator
description	Generate randomized, non-repeating personalities for werewolf or mafia subagents in multi-agent social deduction games. Use when Codex needs to create or refresh agent seat personas, assign one distinct speaking style per seat, produce moderator-ready seat briefs, or keep a full table diverse across public reasoning, pressure style, vote posture, claim timing, and discussion rhythm.

Werewolf Subagent Personality Generator

Use this skill to assign distinct seat personalities before the moderator spawns or briefs werewolf subagents.

Quick Start

1. Draw personalities for the whole table in one batch so uniqueness is enforced in one run.
2. For this repository's 12-player game, run:

python .\.agents\skills\werewolf-subagent-personality-generator\scripts\generate_personalities.py --count 11 --seats P2,P3,P4,P5,P6,P7,P8,P9,P10,P11,P12

3. For each seat, copy only the matching seat block into that seat's private/players/Pn.md, filling ### 本局性格档案 and ### 子代理恢复摘要.
4. After the private file is updated, use that same seat's ### 子代理恢复摘要 as the setup prompt for the first spawn.
5. Keep the draw private. Do not write the full mapping to public player history files.

Workflow

1. Re-roll at the start of each new game unless you intentionally need a reproducible lineup.
2. Pass --seed <value> when you need deterministic replay for debugging, rematches, or transcripts.
3. Pass --format json when another tool needs structured output for writing ### 本局性格档案 and ### 子代理恢复摘要.
4. Write each seat's full persona fields into only that seat's private notebook; do not rely on the initial spawn prompt as the only memory source.
5. Edit references/personality_pool.json only when you need to expand or rebalance the pool.
6. Stop and extend the pool before drawing if the requested seat count exceeds the pool size.

Output Handling

• Treat each generated personality as a stable seat-level speaking style for the whole match.
• Persist each generated personality into that seat's private/players/Pn.md as the authoritative persona record for the match.
• Store both the structured persona fields and a restorable long-form summary so a replacement agent can recover without re-rolling.
• Let the personality affect wording, pacing, pressure style, claim posture, and vote behavior.
• Do not let the personality override rule compliance, role secrecy, or information boundaries.
• Give each subagent only its own generated block; do not leak the full table mapping to every agent.

Respawn / Recovery

• If a seat's subagent crashes, respawn the same seat instead of drawing a new personality.
• On respawn, restore from that seat's private/players/Pn.md, especially ### 本局性格档案, ### 子代理恢复摘要, and ## 私密思考记录.
• Do not treat the original spawn prompt as the only source of truth. The private notebook is the long-lived authority.
• Do not rewrite the persona on respawn unless the moderator explicitly starts a new game and re-rolls the full table.

Resources

• Use scripts/generate_personalities.py for all normal draws. The script loads the pool, samples without replacement, supports optional seeds, and emits moderator-ready seat briefs.
• Read references/personality_pool.json only when editing or auditing the pool. The script already enforces uniqueness and formats the downstream prompt text.