Whetstone

Whetstone is a Claude Code plugin designed to enforce engineering discipline in AI coding agents. The core premise is that AI agents do not lack capability, they lack process. The plugin requires planning before coding, verification before claiming done, root cause analysis before fixes, and code review before merging. It bundles 32 skills, 19 specialized agents, 22 slash commands, an MCP server, and skill injection hooks for subagents. Skills alone handle behavioral discipline, while the full plugin orchestrates entire workflows from brainstorming through code review.

The skills are organized into categories covering the full development lifecycle. Architecture and design skills include an agent native architecture checklist, frontend design with philosophy requirements, and code simplification. Language and framework skills cover React, Node.js, Python, Laravel and PHP, Rust, C, C++, Pine Script, and Tailwind CSS. Infrastructure skills handle PostgreSQL optimization, Terraform and OpenTofu, and Bash scripting. Testing and quality skills address test writing using DAMP principles, two pass code review methodology, receiving feedback, debugging protocols, verification gates, and planning ceremonies. Content and workflow skills cover brainstorming, documentation generation, document review, technical writing, git worktree management, markdown documentation, and file based task tracking. AI and prompting skills include meta prompting patterns, prompt refinement, and session reflection. There is also a multi agent orchestration skill for swarm coordination and dispatch discipline.

The plugin centers on five primary workflow commands. The brainstorm command runs structured requirement exploration producing two to three design approaches with trade offs. The plan command converts requirements into atomic, file specific implementation tasks in vertical slices. The work command executes plans with task tracking and verification gates before marking tasks complete. The review command runs multi agent code review detecting scope drift, validating specifications, and assessing quality, security, and performance. The compound command captures solutions as searchable documentation in a solutions folder so institutional knowledge accumulates rather than getting lost.

The 19 specialized agents divide into review, research, design, and workflow categories. Review agents cover accessibility audits, architecture evaluation, cloud infrastructure, code simplicity, database safety, Python and TypeScript standards, performance analysis, security threats, and specification flow analysis. Research agents gather best practices, excavate git history, mine institutional knowledge, and analyze repository patterns. Design agents handle iterative UI refinement and Figma design compliance. Workflow agents validate bug reproduction, verify deployments, manage infrastructure lifecycle, and resolve PR comments.

Installation for Claude Code uses the plugin marketplace with a few commands. Multi platform support extends the skills to standalone use via a separate ai skills repository that works with 35 or more compatible agents. There is a native Codex plugin with installation scripts, an OpenCode converter, and symlink based deployment for other tools. The skills are format agnostic, written as plain Markdown with token budgets under 1K tokens per skill and a 2K maximum, ensuring they load quickly and do not overwhelm context windows. Activation is automatic based on keyword rich descriptions, file type, and task signals rather than requiring manual toggling.

The distillery directory automates skill generation and improvement. It aggregates top rated community skills, synthesizes overlapping guidance, evaluates effectiveness using an LLM as judge pattern, evolves instructions using DSPy optimization, and mines Claude Code session logs for training data. This means the skill library improves over time rather than remaining static.

Whetstone is released under the MIT license, which means free to use, modify, and adapt without restriction. There is no paid tier, no signup, and no SaaS component. You grab the plugin from GitHub, install it into your Claude Code environment, and start using it immediately. The target audience is engineers using Claude Code for production work who want to reduce the cycle of generating code, finding bugs, and regenerating. It also fits teams building AI assisted development workflows who need a shared baseline of quality checks across PHP, Python, TypeScript, and React codebases.