Spellbook: A Portable Skill Library for AI Programming Assistants

Spellbook is a portable skill library designed for AI programming assistants, aiming to solve the problem of inconsistent work quality across different AI tools (e.g., Claude Code, Codex, Pi) in various projects. It includes 8 workflow skills and 7 agent definitions. Through structured skill modules, design patterns (such as the GAN triplet), multi-harness support, etc., it helps teams establish consistent AI-assisted development standards, covering the entire workflow from requirement analysis to code delivery.

Today, with the popularity of AI-assisted programming tools, developers face the challenge of inconsistent work quality of AI assistants across different projects and toolchains. Although tools like Claude Code and Codex are powerful, their default configurations lack deep optimization for specific scenarios. Spellbook emerged as a portable skill library that encodes best practices into reusable modules, helping teams establish consistent AI-assisted development standards regardless of the AI tools or tech stacks used.

Spellbook contains 8 core workflow skills covering all stages of the software development lifecycle:

• /autopilot: End-to-end delivery, coordinating planning, building, reviewing, and releasing;
• /code-review: Parallel multi-agent review and automatic fixing;
• /investigate: Problem investigation and fixing;
• /groom: Backlog management and requirement grooming;
• /harness: Skill engineering (primitive management, context control);
• /reflect: Conversation review and continuous improvement;
• /research: Technical research and solution evaluation;
• /shape: Converting specifications into structured context data packets.

Its core design concept is the GAN triplet model: Planner (e.g., /shape produces design documents), Builder (e.g., /autopilot executes implementation), and Judge (e.g., /code-review for quality assessment), which improves output quality through iterative cycles. Additionally, a virtual design review committee (incorporating perspectives from four experts: Ousterhout, Carmack, Grug, and Beck) is introduced to avoid single-thinking limitations and balance technical debt with development speed.

Spellbook defines a complete development workflow: backlog.d/ collects requirements → /groom organizes into structured backlogs → /shape produces detailed specifications → /autopilot implements → /code-review checks → release, supporting iterative backtracking. It supports multiple harnesses like Claude Code, Codex, and Pi. Through the bootstrap.sh script, it automatically discovers skills and creates symbolic links to each harness's configuration directory, ensuring that a single skill definition is used consistently across all tools. New harnesses can be extended by adding an adaptation layer.

Spellbook defines best practices for skill development: New skills need to have a SKILL.md file under skills/{name}/, containing metadata (purpose, input/output, dependencies); follow the "traps before instructions" principle (list error cases instead of just instructions) and the "progressive disclosure" architecture (from concise descriptions to in-depth documentation).

The project structure is clear: skills/ stores 8 workflow skills, agents/ stores 7 agents, harnesses/ organizes configurations by tool (including shared general principles), registry.yaml maintains external skill sources, and bootstrap.sh is the entry script. The project remains concise, resists uncontrolled expansion, and uses symbolic links to ensure real-time reflection of changes.

Spellbook provides systematic workflow templates for individual developers, ensures consistency in AI usage for teams, and serves as a knowledge management carrier for organizations. Its practical values include reducing repetitive prompt engineering, improving AI code quality, accelerating new member onboarding, and establishing auditable processes.

The project uses the MIT license and encourages community contributions (adding new skills, improving definitions, expanding harness support). Contributors are required to stress-test skills after model upgrades to maintain robustness and conciseness.