create-twin

name	create-twin
description	Distill any person into a reusable Claude Code skill. Feed it chat logs, docs, emails, git history, and get back a digital twin that thinks, decides, and communicates like them.
user-invocable	true

create-twin

Distill a person into a digital twin skill. Inspired by colleague-skill, adapted for founders, engineers, and operators.

Trigger

User says: "create twin", "distill me", "make a skill of me", "digitize [name]", "create a twin of [name]"

Process

Step 1: Intake (3-5 questions, < 2 min)

Ask these in rapid succession. All except alias are skippable.

1. Alias: What should this twin be called? (e.g., "homen", "sarah-cto")
2. Identity in one sentence: Role, company, domain, stage. (e.g., "Solo founder, agentic AI, pre-seed, banking background")
3. Operating philosophy in one sentence: How they make decisions. (e.g., "Prove first, plan second. Binary gates. Kill criteria before starting.")
4. What they push back on: 2-3 things that trigger them. (e.g., "Scope creep, band-aid fixes, passive AI that waits for instructions")
5. What excites them: 2-3 things they lean into. (e.g., "Self-building systems, compounding memory, speed as emotion")

Step 2: Source Material Collection

Gather from every available source. The more material, the higher fidelity.

Auto-collect (if accessible):

• Git commit messages: git log --author="[name]" --format="%s" | head -100
• CLAUDE.md rules and .claude/rules/*.md
• Memory files in ~/.claude/projects/*/memory/
• SOUL.md, STYLE.md if present
• Docs/ folder strategy docs, PRDs, specs

User-provided (ask for these):

• Chat logs (Slack, Discord, iMessage exports)
• Email samples (especially sent mail — tone, structure, priorities)
• LinkedIn posts or public writing
• Meeting transcripts
• Decision memos or strategy docs
• Code review comments

Connected tools (if MCP connectors available):

• Gmail: read sent mail for tone and priorities
• Slack: read message history for communication patterns
• Google Drive: read strategy docs and memos
• Linear/Jira: read issue comments for delegation style

Step 3: Parallel Analysis

Run two extraction tracks on all collected material:

Track A: Work Skill Extraction For each piece of source material, extract:

• How they decompose problems (top-down? bottom-up? constraint-first?)
• How they evaluate tradeoffs (frameworks used, what they weigh)
• How they structure plans (sequenced steps? parallel tracks? timebox?)
• How they benchmark claims (what evidence do they demand?)
• Output format preferences (tables? bullets? schemas? code?)
• Technical stack and tool preferences
• Quality bar (what's "done" to them?)

Track B: Persona Extraction For each piece of source material, extract:

Layer	Extract	Format
Layer 0: Core Rules	Inviolable behavioral rules	"In situation X, does Y" — NOT adjectives
Layer 1: Identity	Role, domain, stage, background	One paragraph
Layer 2: Expression	Vocabulary, sentence length, connectors, formality, emoji usage, catchphrases. Plus 4-5 "sounds like them" examples.	Concrete patterns
Layer 3: Decision	Priority ranking, push triggers, avoidance triggers, how they say "no", how they handle uncertainty	Situation-behavior pairs
Layer 4: Interpersonal	Delegation style, feedback style, behavior toward superiors/peers/juniors, behavior under pressure	Scenario examples
Layer 5: Boundaries	Things they dislike, refuse, or avoid. What triggers pushback.	Evidence-based

Step 4: Generate Twin Skill

Write the skill to ~/.claude/skills/{alias}-twin/SKILL.md with this structure:

---
name: {alias}-twin
description: Digital twin of {name}. {one-sentence identity}.
trigger: when user says "{alias} mode", "be {alias}", "write as {alias}"
---

# {name} — Digital Twin Skill

## Identity
{Layer 1 output}

## Modes
{3-7 modes extracted from work patterns, each with: when to use, tone, output format}

## Communication Style
{Layer 2 output — voice rules, sentence patterns, vocabulary}

## Decision-Making
{Layer 3 output — frameworks, gates, triggers}

## Work Methodology
{Sequenced steps from Track A analysis}

## Push-Back Patterns
{Layer 5 output — what they interrupt and how}

## Excitement Signals
{What they lean into, from intake + material analysis}

## Calibration Set
### Sounds like {name}
{4-5 example sentences/paragraphs from real material}
### Does NOT sound like {name}
{4-5 anti-examples — common AI patterns they'd reject}

## Layer 0: Inviolable Behavioral Rules
{10-15 situation-behavior pairs, concrete, from Layer 0 extraction}

## Corrections
{Empty section — grows through use}

## Meta
Distilled from: {source list with counts}. Last updated: {date}.

Step 5: Preview + Confirm

Show the user a summary before writing:

• Identity (1 sentence)
• Top 3 modes detected
• Top 5 Layer 0 rules
• 2 "sounds like" examples
• 2 "does not sound like" examples

Ask: "Does this feel right? Anything to correct?"

Apply corrections, then write the file.

Step 6: Evolution

After initial creation, the twin improves through:

Corrections: User says "I wouldn't say that" or "that's not me" → extract [situation, wrong, correct] → add to Corrections section.

New material: User provides more source material → re-run analysis → merge new patterns via incremental update (check for conflicts, prompt user on contradictions).

Calibration: Periodically ask "does this sound like you?" on generated outputs. Track accuracy over time.

Output

The skill file at ~/.claude/skills/{alias}-twin/SKILL.md is immediately usable in any Claude Code session. Trigger with "{alias} mode" or "be {alias}".

Anti-patterns

• Vague adjectives instead of behavioral rules ("decisive" → useless. "When presented with two options and no data, picks the faster-to-validate" → useful.)
• Overfitting to one data source (git commits show work, not personality. Emails show tone, not technical depth. Need multiple sources.)
• Copying colleague-skill's Chinese tech company level mapping (irrelevant for most users — skip it.)
• Generating without preview/confirmation (the user must validate Layer 0 rules before the skill is written.)