퇴사하는 동료의 메시지와 문서로 만든 Distilly Person Profile 스킬

# Distilly Person Profile: Build Plan for a Departing Teammate

## 1. Candidate Source Inventory

**Exported chat threads**
- Slack/Discord/Teams DM and channel exports where they were a primary participant
- Threads where they made a call (approvals, "let's go with X", "no, because…")
- Incident channels and design-review threads

**Docs they authored**
- Design docs, RFCs, ADRs
- Runbooks and on-call notes
- PR descriptions and review comments (often the richest decision trail)
- Meeting notes they wrote

**Interview transcripts**
- Hiring loops they ran (shows calibration and standards)
- Their own onboarding/exit interviews
- Customer or user-research interviews they conducted

**Public talks and writing**
- Conference talks, slides, recordings
- Blog posts, Dev.to/Hacker News/Reddit/Lobste.rs comments
- Open-source PRs and issues on GitHub

**Missing / likely gaps to chase down**
- Verbal decisions never written down (ask for a 30-min brain dump before they leave)
- Their personal notes app / scratchpad exports
- Code review comments on repos you don't have access to
- Any private DMs where real reasoning happened (get consent + export)
- Performance-review self-assessments (HR-sensitive, see §3)
- Recordings of meetings where they were the decision-maker

## 2. Proposed Folder Layout

```
distilly/
├── raw/                          # untouched sources, never edited
│   ├── chat/
│   │   ├── slack-export-2024-01/
│   │   └── threads-curated.md
│   ├── docs/
│   │   ├── rfcs/
│   │   ├── adrs/
│   │   └── runbooks/
│   ├── interviews/
│   │   ├── hiring-loops/
│   │   └── user-research/
│   ├── public/
│   │   ├── talks/
│   │   └── writing/
│   └── MANIFEST.md               # what each file is, date, provenance
├── scrubbed/                     # redacted copies safe to commit
│   └── (mirrors raw/ structure)
├── profile/                      # generated output only
│   ├── role-and-decisions.md
│   ├── voice-and-style.md
│   └── eval-set.md
└── README.md                     # what this is, consent note, retention
```

Keep `raw/` gitignored or in a private store; only `scrubbed/` and `profile/` get committed.

## 3. Files to Flag Before Committing

Scrub or exclude anything matching:
- **Credentials**: `.env`, `secrets.*`, config with API keys, chat pastes containing tokens, `~/.aws`, connection strings
- **Customer PII**: support tickets, user-research transcripts with names/emails, CRM exports, logs with user IDs
- **Private HR content**: performance reviews, comp discussions, PIPs, hiring debriefs with candidate names, exit-interview notes
- **Legal/contract**: anything under NDA, vendor contracts, security incident postmortems with embargo

Add a pre-commit check (e.g., `gitleaks` + a manual `scrubbed/` review) and a `MANIFEST.md` column marking each source `public | internal | restricted`.

## 4. Role-and-Decisions Brief for the Distilly Creator

> **Who they were:** [Name], [role] on [team], [dates]. Owned [scope]. Primary decision-maker for [domains].
>
> **Judgment calls to preserve:**
> - How they scoped ambiguous problems (what they cut first, what they refused to cut)
> - Their default tradeoff posture (ship vs. harden, build vs. buy, monolith vs. service)
> - How they handled disagreement — when they overrode, when they deferred, and the signals they used
> - Their bar for "done" and for "good enough to ship"
> - How they prioritized when everything was urgent
> - What they escalated vs. absorbed
> - Their instincts on risk, rollback, and reversibility
>
> **Voice:** [direct/diplomatic], [terse/expansive], uses [examples/analogies/data]. Avoids [jargon/hedging].
>
> **Do not infer:** anything about comp, performance, or personnel matters.

## 5. Ten Past Decisions for the Eval Set

Pick real, dated decisions with a clear "what they chose and why." Suggested shape:

1. **Architecture**: chose [X] over [Y] for [service] — capture the reasoning and what they rejected.
2. **Build vs. buy**: decided to build [component] in-house rather than adopt [vendor].
3. **Scope cut**: dropped [feature] to hit [deadline] — what they protected instead.
4. **Hiring bar**: rejected/advanced a candidate on [signal] — tests calibration.
5. **Incident call**: during [outage], chose [rollback vs. fix-forward] — tests risk posture.
6. **Tech debt**: deferred vs. paid down [debt] — tests prioritization.
7. **Process**: introduced or killed [process] — tests how they weigh overhead.
8. **Cross-team conflict**: sided with [team A] over [team B] on [issue] — tests disagreement handling.
9. **Tooling/stack**: adopted [TypeScript/Python/LangChain/MCP/RAG approach] for [reason] — tests technical taste.
10. **Escalation**: escalated vs. absorbed [problem] — tests judgment on ownership.

For each, store: the situation, the options, what they chose, their stated reasoning, and 2–3 "acceptable answer" variants so the eval checks reasoning, not exact wording.

Want me to draft the `MANIFEST.md` template and a scrub checklist next?