Gerstep/HumanCompiler

 ██╗  ██╗██╗   ██╗███╗   ███╗ █████╗ ███╗   ██╗
 ██║  ██║██║   ██║████╗ ████║██╔══██╗████╗  ██║
 ███████║██║   ██║██╔████╔██║███████║██╔██╗ ██║
 ██╔══██║██║   ██║██║╚██╔╝██║██╔══██║██║╚██╗██║
 ██║  ██║╚██████╔╝██║ ╚═╝ ██║██║  ██║██║ ╚████║
 ╚═╝  ╚═╝ ╚═════╝ ╚═╝     ╚═╝╚═╝  ╚═╝╚═╝  ╚═══╝

  ██████╗ ██████╗ ███╗   ███╗██████╗ ██╗██╗     ███████╗██████╗
 ██╔════╝██╔═══██╗████╗ ████║██╔══██╗██║██║     ██╔════╝██╔══██╗
 ██║     ██║   ██║██╔████╔██║██████╔╝██║██║     █████╗  ██████╔╝
 ██║     ██║   ██║██║╚██╔╝██║██╔═══╝ ██║██║     ██╔══╝  ██╔══██╗
 ╚██████╗╚██████╔╝██║ ╚═╝ ██║██║     ██║███████╗███████╗██║  ██║
  ╚═════╝ ╚═════╝ ╚═╝     ╚═╝╚═╝     ╚═╝╚══════╝╚══════╝╚═╝  ╚═╝

Compile humans into AI agents.

A Claude Code plugin that conducts deep behavioral interviews, reads work artifacts via MCP tools, and generates installable plugins that authentically embody a person's thinking, communication, and decision-making.

How It Works

                    ┌─────────────────────────────────┐
                    │         /compile-human           │
                    └──────────────┬──────────────────┘
                                   │
                    ┌──────────────▼──────────────────┐
                    │      8-Phase Deep Interview      │
                    │                                  │
                    │  1. Identity & Role               │
                    │  2. Communication Style           │
                    │  3. Decision-Making               │
                    │  4. Domain Expertise              │
                    │  5. Work Patterns                 │
                    │  6. Edge Cases                    │
                    │  7. Artifact Analysis (MCP)       │
                    │  8. Calibration & Corrections     │
                    └──────────────┬──────────────────┘
                                   │
                    ┌──────────────▼──────────────────┐
                    │   MCP Source Analysis             │
                    │                                  │
                    │   Notion ─── docs, meeting notes  │
                    │   Asana ──── tasks, projects      │
                    │   Chrome ─── web artifacts        │
                    │   Local ──── files, transcripts   │
                    └──────────────┬──────────────────┘
                                   │
                    ┌──────────────▼──────────────────┐
                    │     Behavioral Profile (YAML)    │
                    │                                  │
                    │   ~/.human-compiler/<name>/      │
                    │   ├── profile.yaml               │
                    │   ├── phases/                    │
                    │   └── artifacts/                 │
                    └──────────────┬──────────────────┘
                                   │
                    ┌──────────────▼──────────────────┐
                    │    Plugin Generator               │
                    │                                  │
                    │    Profile ──► Handlebars ──► Plugin│
                    └──────────────┬──────────────────┘
                                   │
                    ┌──────────────▼──────────────────┐
                    │   Installable Claude Code Plugin  │
                    │                                  │
                    │   <name>-autonomous  (acts)       │
                    │   <name>-advisory   (advises)     │
                    │   /ask-<name>       (consults)    │
                    └─────────────────────────────────┘

Quick Start

# 1. Install the plugin
claude --plugin-dir /path/to/HumanCompiler

# 2. Start the interview
/compile-human

# 3. After all 8 phases, install the generated agent
claude --plugin-dir ~/.human-compiler/<name>/output-plugin/<name>-agent

Features

Deep Behavioral Interview

• 8 structured phases covering identity, communication, decisions, expertise, workflows, edge cases
• 10-20 adaptive questions per phase with follow-up probes
• Concrete examples over abstractions — captures real behavior, not aspirations

MCP-Powered Artifact Analysis

• Reads Notion pages, meeting notes, strategy docs for communication style
• Analyzes Asana tasks and projects for work patterns
• Cross-references artifacts with interview answers for authenticity

Progressive & Resumable

• Saves after every phase — never lose progress
• /compile-human resume picks up where you left off
• /compile-human status shows progress at a glance

Configurable Autonomy

• Autonomous mode (acceptEdits): Agent acts, decides, and communicates as the person
• Advisory mode (plan, read-only): Agent recommends in the person's style without acting

Marketplace-Ready Output

• Generates a complete Claude Code plugin with manifest, agents, skills, CLAUDE.md
• Install via --plugin-dir or distribute through plugin marketplace

Output Plugin Structure

<name>-agent/
├── .claude-plugin/
│   └── plugin.json              # Plugin manifest
├── agents/
│   ├── <name>-autonomous.md     # Full autonomy agent
│   └── <name>-advisory.md       # Advisory-only agent
├── skills/
│   ├── ask-<name>/
│   │   └── SKILL.md             # "What would <name> do?"
│   └── <domain>-advice-<name>/
│       └── SKILL.md             # Expertise-specific skill
└── CLAUDE.md                    # Agent documentation

Commands

Architecture

HumanCompiler/
├── .claude-plugin/plugin.json         # Plugin manifest
├── skills/compile-human/
│   ├── SKILL.md                       # Orchestrator skill
│   ├── interview-guide.md             # Interview methodology
│   └── phase-instructions/            # Per-phase question banks
│       ├── 01-identity.md
│       ├── 02-communication.md
│       ├── 03-decision-making.md
│       ├── 04-domain-expertise.md
│       ├── 05-work-patterns.md
│       ├── 06-edge-cases.md
│       ├── 07-artifact-analysis.md
│       └── 08-calibration.md
├── agents/interviewer.md              # Interview agent persona
├── scripts/
│   ├── profile-manager.ts             # Profile CRUD operations
│   ├── profile-manager-cli.ts         # CLI wrapper
│   ├── generate-plugin.ts             # Profile → Plugin generator
│   └── templates/                     # Handlebars templates
│       ├── plugin.json.hbs
│       ├── agent-autonomous.hbs
│       ├── agent-advisory.hbs
│       ├── skill-ask.hbs
│       ├── skill-task.hbs
│       └── claude-md.hbs
├── tests/                             # Full test suite
│   ├── profile-manager.test.ts
│   ├── template-render.test.ts
│   ├── generate-plugin.test.ts
│   └── e2e-interview-flow.test.ts
├── package.json
└── CLAUDE.md

Profile Schema

The behavioral profile captures:

Development

# Install dependencies
bun install

# Run tests
bun test

# Run specific test suite
bun test profile-manager
bun test template-render
bun test generate-plugin
bun test e2e

# Generate a plugin from a profile
bun run scripts/generate-plugin.ts ~/.human-compiler/<name>/profile.yaml

How the Interview Works

Each phase follows a consistent pattern:

1. Read phase instructions — load the question bank and MCP guidance
2. Query MCP sources — search Notion, Asana, etc. for relevant artifacts
3. Conduct adaptive interview — ask questions, follow interesting threads
4. Cross-reference artifacts — validate answers against real work products
5. Record structured data — update the YAML profile
6. Save raw transcript — preserve the full conversation for reference
7. Summarize and confirm — present findings, ask for corrections
8. Transition — move to the next phase

The final calibration phase presents the complete profile for review and correction, ensuring the compiled agent is authentically representative.

License

MIT

{
  "name": "HumanCompiler",
  "owner": {
    "name": "Gerstep"
  },
  "metadata": {
    "description": "Compile humans into AI agents through deep behavioral interviews"
  },
  "plugins": [
    {
      "name": "compile-human",
      "source": "./",
      "description": "Conduct deep behavioral interviews and generate installable Claude Code plugins that embody a person's decision-making, communication style, and expertise.",
      "version": "0.1.0",
      "homepage": "https://github.com/Gerstep/HumanCompiler",
      "repository": "https://github.com/Gerstep/HumanCompiler",
      "license": "MIT",
      "keywords": [
        "agent",
        "interview",
        "digital-twin",
        "compiler",
        "cowork"
      ]
    }
  ]
}