How it works

XeroHack runs a four-stage pipeline. Each stage narrows the design space so the output is opinionated rather than generic — the opinion comes from the directory.

1. Discovery

You paste, type, or drop your idea onto a single open chat surface. A deep-thinking model (anthropic/claude-opus-4.7 by default, swappable via OpenRouter) expands the input into a structured brief:

{
  "problem": "...",
  "audience": "B2B | D2C",
  "surface": "web | mobile | cli | hybrid",
  "must_haves": [],
  "nice_to_haves": [],
  "constraints": [],
  "risks": []
}

2. Interview

The engine proposes 5–7 architecture questions, presented one at a time. Each question is:

• Long-form with context, not multiple-choice.
• Ordered so each answer narrows the next — UI framework before component library before charting library.
• In-line with the trade-offs: not "pick one," but "here's why each side wins and what you give up."

Vendor and tool questions surface options that exist in the directory first, and apply discount codes automatically when a vendor has one.

3. Configuration

The engine takes your answers and emits a complete scaffold to a virtual filesystem, then ships it as a .zip and — optionally — pushes it to a fresh GitHub repo. Every file carries the XeroHack breadcrumb and license header. See Anatomy of an output for the full tree.

4. Execution

After download or push, the engine surfaces everything you need to start:

• The exact install command sequence for your chosen harness.
• The starter prompt to paste into Claude Code, Cursor, or Codex.
• The README content, ready to publish.

The plan is structured as waves → sprints → epics, designed for swarm execution with worktrees so parallel agents don't conflict.