# Adapter: nuwa-skill → claw-boss-distiller

We vendor `upstream/` (= [`alchaincyf/nuwa-skill`](https://github.com/alchaincyf/nuwa-skill), MIT) and adapt it for our specific job: distill a private boss from their own private email corpus.

## What we keep from nuwa

- **5-layer output structure** (Expression DNA / Mental Models / Decision Heuristics / Anti-Patterns / Honest Boundaries) — directly applicable
- **Triple verification** (cross-domain / generative / exclusive) — same logic, prevents random one-off comments from being elevated to "the boss's mental model"
- **Quality self-check** (`references/extraction-framework.md` § 六) — copy verbatim
- **6-bucket collection paradigm** — but rewired (see below)
- **Phase 4 validation** (sanity / edge / voice) — keep as-is, just substitute the test set
- **Honest Boundaries discipline** — never invent, mark "信息不足"

## What we replace

| nuwa default | Atlas adaptation | Why |
|---|---|---|
| Public figure (Munger, Jobs, etc.) | Private boss at one client | The whole point — we cannot search "what does Boss X think about Y" online; we have only their email corpus |
| 6 web-search agents (books / podcasts / Twitter / external views / decisions / timeline) | 6 email-bucket agents (see below) | No public corpus; the email corpus *is* the data |
| Output as standalone perspective skill (`xxx-perspective`) | Output as `boss_skill.md` consumed by Atlas's other 5 skills | We don't role-play the boss — we cite their rules |
| Role-play activation (`use Munger to analyze...`) | **No role-play**. `boss_skill.md` is a rule library other skills read — never spoken in first person | We are an observer/advisor for the boss, not an impersonator |
| Public release | Private to the client deployment | Privacy-critical; never leaves client network |

## What we add

| Addition | Reason |
|---|---|
| **Boss confirmation gate** between draft and live | Boss owns their own model; we propose, boss disposes |
| **Status field** on every rule (`draft` / `confirmed` / `retired`) | Audit trail; rule lifecycle is observable |
| **Source email IDs** per rule | Boss can drill from rule → "show me the 7 emails this came from" |
| **Quarterly diff workflow** | New emails arrive constantly; rules drift; quarterly refresh + diff for boss |
| **`boss_skill.history/` archive** | Every confirmed version preserved for compare |

## Six Email Buckets (replaces nuwa's 6 web agents)

| Bucket | Scope | What we look for |
|---|---|---|
| **B1: Outgoing-by-intent** | All boss-sent emails grouped by intent (催办/决策/转交/否决/表扬) | How boss expresses each intent; boundary phrases; tells |
| **B2: Mid-thread interventions** | Threads where boss jumped in mid-conversation | High-signal moments — boss only intervenes when something matters |
| **B3: Short replies (< 50 words)** | Compressed-mode boss thinking | Reveals what boss treats as obvious; the unstated frame |
| **B4: Long replies (> 200 words)** | Reasoning-mode boss thinking | Reveals full chain of thought; the explicit frame |
| **B5: Pressure threads** | Threads with customer-side complaints / urgent escalations | How boss responds under stress; non-negotiables surface |
| **B6: Closing threads** | Threads boss explicitly closed (verbal verdict, deliverable accepted, project killed) | Decision criteria + closing language |

This replaces nuwa's `references/research/01–06.md` collection schema. The aggregate output still lives at `references/research/` but each `0X-xxx.md` is now an email-bucket index, not a web-search dump.

## Output Format Diff

nuwa output has `## 角色扮演规则` and `## 身份卡` sections — **we delete both.** Atlas's `boss_skill.md` is a rule library, not an impersonation.

We also delete `## 智识谱系` (intellectual lineage) — interesting for public figures, irrelevant for a private boss.

We keep, in order:
1. `## L1 — Expression DNA`
2. `## L2 — Mental Models`
3. `## L3 — Decision Heuristics`
4. `## L4 — Anti-Patterns`
5. `## L5 — Honest Boundaries`
6. `## Source Index` (which buckets / how many emails fed this version)
7. `## Audit Log` (status changes, last reviewed, last fired counts)

## Mapping Table for the Implementor

| nuwa file | Atlas equivalent | Status |
|---|---|---|
| `upstream/SKILL.md` | `claw-boss-distiller/SKILL.md` (already written) | Done |
| `upstream/references/extraction-framework.md` | Reused as-is via `upstream/` | Linked |
| `upstream/references/skill-template.md` | `claw-boss-distiller/boss_skill.template.md` (delete role-play & 智识谱系 sections) | Done — see `boss_skill.seed.md` for filled example |
| `upstream/scripts/` | Possibly reused for cleanup utilities; the rest don't apply (we don't download YouTube subtitles) | Skip in V0 |
| `upstream/examples/*-perspective/` | Not applicable (no public figures); keep for reference only | Reference only |

## Two-Pass Run Pattern (V0 default)

### Pass 1 — Cold Start (W1 of new deployment)

- Input: 6 months of boss outgoing email
- 6 buckets populated in parallel
- Triple verification yields ~20 candidate rules
- Output: `boss_skill.md` v0 with **all rules `status: draft`**
- Boss audit session in W2 → bumps confirmed rules to `status: confirmed`

### Pass N — Quarterly Refresh

- Input: 90 days of new outgoing email + override log from past 90 days
- Same 6-bucket pipeline
- Diff vs current `boss_skill.md`:
  - **Add candidates** — patterns now seen ≥ 3 times that weren't there before
  - **Revise candidates** — confirmed rules whose evidence has shifted
  - **Retire candidates** — confirmed rules with 0 fires in 90 days OR overridden ≥ 2 times
- Boss reviews diff, applies, archives prior version under `boss_skill.history/YYYY-Q*.md`

## Why we don't simply use nuwa unchanged

- nuwa's role-play default would have Atlas speak *as* the boss — wrong product. We surface signals; we don't impersonate.
- nuwa's web-search agents would crash here — there's no web corpus for a private boss.
- nuwa expects the output to be the deliverable; we expect the output to be a rule library other skills consume.

But the **methodology** — how to distinguish a real mental model from random noise — is exactly what nuwa does well. So we vendor + adapt.