assistant-claw/atlas/skills/claw-boss-distiller/SKILL.md

---
name: claw-boss-distiller
description: Distill the boss's mental model from outgoing email corpus into a 5-layer boss_skill.md (Expression DNA / Mental Models / Decision Heuristics / Anti-Patterns / Honest Boundaries). Derived from nuwa-skill (MIT). Drafts only — boss confirms every change.
version: 0.1.0
author: Moments / Atlas team (adapted from alchaincyf/nuwa-skill, MIT)
license: MIT
metadata:
  hermes:
    category: atlas
    tags: [mental-model, distillation, nuwa, boss-os, atlas]
    related_skills: [claw-email-parser, claw-report-writer]
    upstream_dependency: alchaincyf/nuwa-skill
    boundaries:
      - all_new_rules_default_to_status_draft
      - never_apply_changes_without_explicit_boss_confirm
---

# claw-boss-distiller

## Purpose

Capture and codify how the boss thinks, decides, and writes — so the rest of Atlas's skills judge the way the boss would judge, and the Brief reads in the boss's voice.

This is the heart of the "老板 OS" / "女娲 skill" promise. Without this, Atlas is generic; with this, Atlas becomes a personal layer.

## Heritage

Adapted from [`alchaincyf/nuwa-skill`](https://github.com/alchaincyf/nuwa-skill) (MIT). The original distills public figures from open data; we replace the corpus with the boss's private outgoing email and add boss-confirmation gates.

We vendor the upstream `references/extraction-framework.md` and adapt:
- Data source: outgoing email (last 6 months by default)
- Output: `boss_skill.md` (5 layers, in the boss's primary writing language)
- Audit gate: every produced rule starts as `status: draft` and requires boss confirmation to become `status: confirmed`

## Inputs

- Boss outgoing email corpus (from `state/extracted/`, filtered `from = boss email`)
- Current `boss_skill.md` (for diff)
- Optional: 3–5 sample threads where the boss made a notable decision (boss can hand-pick for first run)

## Outputs

- `boss_skill.md` (proposed new version)
- `boss_skill.diff.md` (vs current)
- `boss_skill.evidence/` directory (JSON per rule with source email IDs and quoted excerpts)

## 5-Layer Output Structure

```markdown
# boss_skill.md

## Layer 1 — Expression DNA
- Tone: …
- Sentence length distribution: …
- Signature phrases: …
- Greeting / closing patterns: …
- Few-shot examples (5 representative emails)

## Layer 2 — Mental Models (3–7 frameworks the boss reaches for)
- "先抓主要矛盾" — applied when …
- "凡事预则立" — applied when …
- (etc. — each with 2–3 example threads as evidence)

## Layer 3 — Decision Heuristics (~50 if-then rules, the L1 ruleset)
- R-01: 超过 5 个工作日无回复的项目 → Stalled  [confidence: high; evidence: 7 instances]
- R-02: 客户方 2 次以上主动催办 → 升级红色风险
- … (each rule: id, condition, action, confidence, evidence count, status, last-fired)

## Layer 4 — Anti-Patterns (what the boss explicitly avoids)
- AP-01: 不接受"我以为是别人在跟"作为停滞理由
- AP-02: 不在客户面前批评内部员工
- … (each: id, rule, evidence)

## Layer 5 — Honest Boundaries (what Atlas should NOT do)
- HB-01: 不替老板回复客户邮件 (V0)
- HB-02: 不给员工打 A/B/C 评级
- HB-03: 跨客户数据隔离，不在 A 客户报告里出现 B 客户信息
- … (these come from SOUL.md + boss-confirmed extensions)
```

## 4-Stage Workflow (from nuwa)

### Stage 1 — Multi-Path Collection
6 collection paths (adapted to email context):
1. Outgoing emails by intent: 催办 / 决策 / 转交 / 否决 (4 buckets)
2. Threads where boss intervened mid-conversation (high-signal moments)
3. Email replies under 50 words (压缩态思维)
4. Email replies over 200 words (完整推理)
5. Threads with customer-side complaints — how boss responds under pressure
6. Threads marking project completion — how boss closes things

### Stage 2 — Triple Verification
Every candidate rule must pass:
1. **Cross-domain occurrence** — pattern appears across ≥ 3 different projects/contexts
2. **Predictive power** — given a new thread, the rule predicts the boss's action correctly
3. **Exclusivity** — pattern is distinctive to this boss vs generic corporate norms

Patterns failing any check become Stage-1 evidence only, not rules.

### Stage 3 — Skill Construction
Assemble the 5-layer document. Draft status on every rule.

### Stage 4 — Quality Validation
- Run on 10 held-out historical threads — does the rule predict boss's actual reply?
- Run on 5 currently-undecided threads — produce a recommendation, present alongside boss's eventual decision
- Compute hit rate; rules with hit rate < 70% downgraded or retired

## Boss Confirmation Loop

1. Distiller produces `boss_skill.md` draft + diff
2. Atlas sends boss a "Quarterly Skill Review" email with the diff and a per-rule action checklist
3. Boss replies with acceptances / edits / rejections (parsed via simple structured-reply format)
4. Atlas applies confirmed changes, archives prior version under `boss_skill.history/YYYY-Q*.md`

## Failure Modes

| Failure | Behavior |
|---------|---------|
| Insufficient data (< 200 outgoing emails) | Halt distillation, surface to boss with explanation |
| Stage-2 verification yields < 10 rules | Distillation marked "preliminary", boss audit deferred |
| Boss does not respond to review within 30 days | Atlas sends one reminder; if still no response, freeze current `boss_skill.md`, log "review skipped" |

## Sample Distillation Run

```
Input: 1247 outgoing emails over past 90 days
Stage 1: 412 candidate signals across 6 paths
Stage 2: 73 pass cross-domain, 51 pass predictive, 38 pass exclusivity
Stage 3: 38 rules organized into 5 layers
Stage 4: hit rate 81% (well above 70% threshold)
Output: boss_skill.md v2 (12 new rules, 8 revised, 3 retired vs v1)
Boss action: review meeting scheduled
```