Zeroshot

Overview

Zeroshot automates engineering tasks by pointing Claude Code at a GitHub issue and letting multiple isolated agents check each other's work. The key insight: a validator that didn't write the code has no stake in defending it.

Key Features

• Multi-agent validation - Separate agents write and validate code, eliminating self-serving bias
• Crash recovery - zeroshot resume restores multi-hour tasks after interruptions
• Daemon mode - Run overnight batch processing with -d flag
• GitHub integration - Point directly at issues for end-to-end automation

Code Snippets

Installation

npm install -g @covibes/zeroshot

Requires Node 18+, Claude Code CLI, and GitHub CLI.

Basic Usage

# Point at a GitHub issue
zeroshot run --issue 123

# Resume after crash
zeroshot resume

# Daemon mode for overnight runs
zeroshot run -d --issue 123

When to Use Zeroshot

Zeroshot excels with well-defined tasks where you can describe what "done" looks like:

• Known unknowns (implementation unclear) - The planner figures it out
• Clear acceptance criteria - Validators can verify completion

For unknown unknowns, use single-agent Claude Code for exploration first, then return with a defined task.

Connections

• ralph-wiggum-technique-guide - Zeroshot productionizes the Ralph loop concept, adding multi-agent validation to the simple bash loop pattern
• building-effective-agents - Implements the orchestrator-workers pattern from Anthropic's agent design guide
• 12-factor-agents - Embodies factors like "small focused agents" and "own your control flow" through its multi-agent architecture