AI Dev Tasks

AI Dev Tasks is a structured task management system for AI development agents created by Ryan Carson (GitHub: snarktank). It provides a simple, proven framework for breaking down features into manageable tasks that AI agents can execute systematically.

Core Concept

Rather than giving AI agents monolithic feature requests (“build a login system”), AI Dev Tasks breaks work into small, discrete, machine-verifiable tasks. This structured approach dramatically improves success rates and reduces iteration cycles.

The Three-File System

1. create-prd.md

Guides AI in generating comprehensive Product Requirement Documents (PRDs) from feature descriptions. The PRD contains:

• Feature overview and goals
• User stories and use cases
• Technical requirements and constraints
• Acceptance criteria
• Edge cases and error handling

2. generate-tasks.md

Converts a PRD markdown file into prd.json with granular, machine-verifiable tasks. Key characteristics:

• Generates 8-15 small tasks (not 3-5 large ones)
• Each acceptance criterion is boolean pass/fail
• Investigation and implementation are separate tasks
• Browser tests use structured commands
• Each task is independently verifiable

3. process-task-list.md

Manages task-by-task execution with approval checkpoints:

• Executes one task at a time
• Validates task completion
• Gets human approval before moving forward
• Maintains clear progress tracking
• Handles failures and retries

Task Structure

Each task in the system includes:

• Title: Clear, specific task name
• Description: What needs to be done and why
• Acceptance criteria: How to verify completion
• Testing approach: How to validate the work
• Dependencies: What other tasks must complete first

Tasks are designed to be small enough to complete in 30-60 minutes, but meaningful enough to make progress.

Workflow

Feature Description  
    ↓  
create-prd.md (AI generates PRD)  
    ↓  
generate-tasks.md (AI breaks into tasks)  
    ↓  
process-task-list.md (AI executes, human reviews)  
    ↓  
Completed Feature  

Key Principles

Clear Context

• Provide AI agents with structured, explicit requirements
• Define success criteria unambiguously
• Break complexity into understandable pieces

Machine-Verifiable

• Each task has clear pass/fail criteria
• Tests are executable and deterministic
• Progress is measurable and auditable

Human in the Loop

• Human review at checkpoints
• Approval gates maintain quality
• Escalation for complex decisions

Compound Learning

• Successful task patterns are documented
• Failures inform future task design
• System improves with repetition

Advantages Over Traditional Approaches

Aspect	Traditional	AI Dev Tasks
Request size	Large, vague	Small, precise
Verification	Manual testing	Automated, machine-verifiable
Learning	Per-project	Systematic, reusable
Success rate	40-60%	80-90%+
Iteration	Many cycles	Fewer refinements
Documentation	Ad-hoc	Built-in, structured

Integration Points

Works seamlessly with:

• Anthropic Claude and Claude Code
• Amp CLI
• Ralph (autonomous agent loop)
• GitHub for version control and PRs
• Custom CI/CD pipelines

Real-World Example

Feature: Add user authentication

Without AI Dev Tasks:

• “Add user authentication system”
• Agent confused about scope
• Multiple iterations needed
• Inconsistent implementation

With AI Dev Tasks:

1. Generate PRD (auth flow, tech stack, security requirements)
2. Break into tasks:
   • Database schema for users
   • Hash password utility function
   • Login endpoint
   • Session management
   • Error handling tests
   • … 12 more specific tasks
3. Execute one task at a time
4. Human approves progress
5. Feature ships with clear audit trail

Stats

• Stars: 7.4k+
• Forks: 1.7k+
• Language: Markdown + Prompts
• Creator: Ryan Carson

Best Practices

Task Breakdown

• Each task should be completable in 30-60 minutes
• Investigation and implementation are separate
• Testing is explicit and part of the task
• Dependencies are explicitly stated

Success Criteria

• Use boolean pass/fail acceptance criteria
• Include examples of passing and failing cases
• Make criteria machine-verifiable
• Avoid subjective language

Human Review

• Review after critical tasks complete
• Provide clear feedback for refinement
• Document patterns in AGENTS.md or similar
• Update task templates based on learnings

Tools & Platforms Supported

• Cursor with AI agents
• Claude Code with Anthropic API
• Windsurf and other AI coding tools
• Amp CLI for Amp-based agents
• Custom implementations with any LLM

Common Pitfalls to Avoid

• Too large tasks: Break into smaller pieces
• Vague criteria: Make acceptance criteria specific
• Missing context: Provide codebase and design patterns
• No dependencies: Explicitly state task ordering
• No rollback plan: Include error handling tasks

Philosophy

AI Dev Tasks embodies the principle that structure enables autonomy:

• Clear structure allows agents to operate independently
• Well-defined tasks enable reliable automation
• Systematic breakdown compounds learning
• Each completed task informs the next

---

Last updated: January 2025
Confidence: High (open source project documentation)
Practical relevance: 7.4k+ stars, widely adopted pattern
Related: Ryan Carson, Ralph, Kieran Klaassen, Compound Engineering