distill

---
name: distill
description: >-
  Generate a celebrity/persona distillation skill by scraping their X (Twitter)
  account and analyzing their thinking patterns, communication style, and mental models.
allowed-tools: Bash, Read, Write, Edit, Glob, Grep, WebSearch
user-invocable: true
emoji: "🧬"
---

# Distill — Celebrity Persona Distillation

Scrape any X (Twitter) account using the Dokobot CLI, analyze the person's tweets,
and generate a structured persona SKILL.md that captures their thinking patterns,
communication style, mental models, and decision heuristics.

Works with any AI agent that can run shell commands: Claude Code, Cursor, Windsurf, Codex, OpenClaw, etc.

**Prerequisites**: [Dokobot CLI](https://dokobot.ai) (`npm i -g dokobot`), Chrome with Dokobot extension, native bridge (`npx dokobot install-bridge`).

**Usage**: `/distill <x_handle_or_url> [--screens N] [--output PATH] [--lang zh|en]`

Arguments: $ARGUMENTS

## Parameters

- `x_handle_or_url` (required): X handle (e.g. `@elonmusk` or `elonmusk`) or full URL (e.g. `https://x.com/elonmusk`)
- `--screens N`: Number of screens to scroll on the profile page (default: 10, more = more tweets collected)
- `--output PATH`: Output directory for the generated SKILL.md (default: current working directory)
- `--lang zh|en`: Output language — `zh` for Chinese, `en` for English (default: auto-detect from the person's primary language)

## Workflow

Follow these steps **exactly in order**. Do NOT skip steps or combine them.

### Step 1: Parse Input

Extract the X handle from the arguments:
- If a URL like `https://x.com/username` or `https://twitter.com/username` → extract `username`
- If `@username` → extract `username`
- If bare `username` → use as-is

Set defaults:
- `screens` = 10 (or from `--screens`)
- `output` = current working directory (or from `--output`)
- `lang` = auto (or from `--lang`)

### Step 2: Scrape X Profile

```bash
npx dokobot doko read "https://x.com/<USERNAME>" --local --screens <SCREENS> --timeout 600 > /tmp/distill-raw-<USERNAME>.txt
```

### Step 3: Continue Scrolling (if needed)

If the output contains `Session:` with an ID, continue reading:

```bash
npx dokobot doko read "https://x.com/<USERNAME>" --local --session-id <SESSION_ID> --screens 10 --timeout 600 >> /tmp/distill-raw-<USERNAME>.txt
```

Repeat up to 3 times or until you have enough material (~50+ tweets).

### Step 4: Save Tweets to Separate File

Extract all tweets by this person from the raw data and save to `<OUTPUT>/<username>-skill/tweets.md`:

```markdown
# @<username> Tweets

> Source: https://x.com/<username>
> Scraped: <ISO_DATE>
> Screens: <N>

## Original Tweets

### <N>. <date>
<tweet text>

Likes: <N> | Reposts: <N> | Views: <N>

---

## Retweets & Engagement

List the accounts/tweets this person retweeted or quote-tweeted, grouped by theme.
```

Only include tweets authored by `@<USERNAME>` in "Original Tweets". Retweeted content goes in "Retweets & Engagement".

### Step 5: Web Search for Context (optional)

Append results to `tweets.md` under a `## Background Context` section.

```bash
npx dokobot doko read "https://en.wikipedia.org/wiki/<PERSON_NAME>" --local --screens 5 --timeout 300
```

### Step 6: Analyze and Distill

From `tweets.md`, extract:

- **Profile Summary**: name, handle, bio, domain, follower count, posting frequency
- **Core Mental Models** (3-7): recurring thinking frameworks with example tweets
- **Decision Heuristics** (3-8): rules of thumb with application context
- **Expression DNA**: sentence style, vocabulary, tone, rhetorical devices, quirks
- **Core Topics**: themes with approximate frequency
- **Signature Quotes**: 5-10 most representative

### Step 7: Generate SKILL.md

Generate `<OUTPUT>/<username>-skill/SKILL.md`:

```markdown
---
name: <username>-persona
description: >-
  AI persona capturing the thinking patterns, decision frameworks, and communication
  style of <Full Name> (@<username>). Distilled from their X/Twitter presence.
emoji: "🎭"
metadata:
  author: dokobot-distill
  version: "1.0.0"
  source: "https://x.com/<username>"
  distilled_at: "<ISO_DATE>"
  tweets_analyzed: <COUNT>
---

# <Full Name> — AI Persona

<One paragraph bio>

## Activation Triggers

- "<Name>会怎么看" / "What would <Name> think"
- "用<Name>的视角" / "From <Name>'s perspective"
- "切换到<Name>模式" / "Switch to <Name> mode"

## Source Material

Raw tweets and engagement data are in [tweets.md](tweets.md).

## Core Mental Models

### 1. <Model Name>
<Description — reference tweets by number, e.g. "see tweet #3">

(NO inline tweet quotes — reference tweets.md)

## Decision Heuristics

| # | Heuristic | Rule | When to Apply |
|---|-----------|------|---------------|
| 1 | <Name> | <Rule> | <Context> |

## Expression DNA

### Voice Characteristics
- **Sentence style**: <description>
- **Vocabulary**: <description>
- **Tone**: <description>
- **Rhetorical devices**: <description>
- **Quirks**: <description>

### Communication Rules
1. <Rule 1>
2. <Rule 2>

## Core Topics
<Bulleted list with estimated frequency>

## Operational Protocol

1. **Read** tweets.md to refresh on this persona's voice
2. **Classify** the user's question
3. **Select** relevant mental models
4. **Draft** response applying Expression DNA rules
5. **Verify** authenticity against tweets.md
6. Search for current data if factual claims are needed
```

### Step 8: Report to User

1. Show the output file path
2. Brief summary (tweets analyzed, key models found)
3. Suggest how to use: add SKILL.md to your agent's instructions

## Error Handling

- If connection fails: ensure Chrome is open with Dokobot extension, and native bridge is installed (`npx dokobot install-bridge`)
- If no tweets returned: check if the handle is correct or if the account is private