Getting Started¶
Prerequisites¶
- Claude Code installed and authenticated.
Install¶
That's it. No Python, no API keys, no environment variables. Your existing Claude Code subscription handles all LLM calls.
First Run¶
Run the full 4-phase workflow:
Or run individual phases:
| Command | Phase | What it does |
|---|---|---|
/haytham:validate |
WHY | Market research, competitor analysis, GO/NO-GO/PIVOT verdict |
/haytham:specify |
WHAT | MVP scope, capability model, system traits |
/haytham:design |
HOW | Build-vs-buy analysis, architecture decisions |
/haytham:plan |
SPECS | OpenSpec with SHALL requirements and Gherkin scenarios |
What to Expect¶
Each phase runs specialist agents, writes structured output to .haytham/session/, and presents a gate decision for your review. Nothing proceeds without your approval.
.haytham/
project.yaml # Your startup idea
session/
phase-1-why/ # Validation findings and verdict
phase-2-what/ # MVP scope and capabilities
phase-3-how/ # Architecture decisions
phase-4-specs/ # Implementation-ready OpenSpec
A full 4-phase run takes approximately 15-20 minutes.
Local Development¶
To modify agents or commands, clone the repo and work locally:
Modifying Agents¶
Agent prompts are markdown files in agents/. Edit any file, then reload the plugin to pick up changes. Each agent file contains a system prompt with frontmatter for model selection and tool permissions.
Adding a New Agent¶
- Create
agents/your-agent.mdwith appropriate frontmatter and system prompt - Reference it from the relevant command in
commands/
Adding a New Command¶
- Create
commands/your-command.md - The command becomes available as
/haytham:your-command
Testing Changes¶
Run the plugin against a test idea:
Check the output files in .haytham/session/ to verify agent behavior.