通过 SOUL.md 精调助手人格（先灰度、再全量）
解决“回复太官话/太模板化”问题：用明确风格规则重写 SOUL.md，并通过小样本回测控制幽默/直率边界。
XDiscovered 2026-02-13Author @steipete
Prerequisites
- Workspace uses `SOUL.md` as active persona file and can be version-controlled.
- You have 10-20 representative conversation prompts for A/B tone evaluation.
Steps
- Draft persona deltas in `SOUL.md`: remove formulaic openers, define brevity baseline, and add safe directness rules.
- Run A/B replay on historical prompts and compare acceptance rate, user correction frequency, and safety-boundary violations.
- Keep improvements that raise clarity without increasing risky phrasing; revert any rule that causes over-aggressive tone.
- Publish final SOUL.md with changelog note so future sessions know why style changed.
Commands
git diff SOUL.mdgit add SOUL.mdgit commit -m 'chore: refine assistant voice rules'Verify
New replies become shorter and less templated while still passing safety/style checks in sampled conversations.
Caveats
- Personality upgrades must not override safety policies or channel etiquette constraints.
- Single viral prompt may not generalize to all user contexts（需验证）。
Source attribution
This tip is aggregated from community/public sources and preserved with attribution.
Open original source ↗