GitHub - Gerstep/HumanCompiler: Compile humans into AI agents — a Claude Code plugin that conducts deep behavioral interviews and generates installable agent plugins

5 min read Original article ↗ 
 ██╗  ██╗██╗   ██╗███╗   ███╗ █████╗ ███╗   ██╗
 ██║  ██║██║   ██║████╗ ████║██╔══██╗████╗  ██║
 ███████║██║   ██║██╔████╔██║███████║██╔██╗ ██║
 ██╔══██║██║   ██║██║╚██╔╝██║██╔══██║██║╚██╗██║
 ██║  ██║╚██████╔╝██║ ╚═╝ ██║██║  ██║██║ ╚████║
 ╚═╝  ╚═╝ ╚═════╝ ╚═╝     ╚═╝╚═╝  ╚═╝╚═╝  ╚═══╝

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

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

Command Description
/compile-human Start a new interview
/compile-human resume Resume an interrupted interview
/compile-human status Check progress on all profiles
/compile-human generate Force-generate plugin from current profile

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:

Section What it records
Identity Role, org, team, responsibilities, goals
Communication Writing style, tone spectrum, patterns, vocabulary
Decision-Making Framework, prioritization, uncertainty handling, tradeoffs
Expertise Domains (with depth), technical skills, industry knowledge
Work Patterns Daily routine, tools, collaboration, meetings, task management
Edge Cases Conflict resolution, ambiguity, failure response, scenarios
Artifacts Analyzed documents, cross-artifact patterns
Calibration Corrections, additions, confidence score

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