aiforging

AI Forging

A structured AI-assisted development framework for producing robust, maintainable codebases. Test-first, pattern-driven refactoring, domain-driven architecture — shipped as a Claude Code plugin.

AI Forging is a counterpoint to "vibe coding." The thesis: AI is incredibly powerful at generating code, but without structure, every feature shipped makes the codebase worse. AI Forging provides that structure as three stages of a metallurgical forge: Fire → Hammer → Tempering.

• Fire — AI-powered TDD. Tests are written first and capture intent before a single line of implementation exists. Red, Green, Refactor. Uses the superpowers plugin's test-driven-development skill.
• Hammer — Automated pattern-driven refactoring. The hammer-refactor skill dispatches one fresh-context subagent per pattern or anti-pattern file against the session's changed files, in parallel. Built on top of superpowers:subagent-driven-development.
• Tempering — Scalable knowledge capture. The capture-pattern skill watches for corrective moments during interactive code review and offers to persist the lesson as a new .md file in the pattern library. One correction, one file. Adding the 50th pattern costs no more than the 5th, because every pattern lives in its own file and gets its own subagent on every Hammer pass.

Each iteration through the cycle leaves the codebase stronger than it started.

Day-to-day usage

Once your workspace and targets are set up, these are the commands you'll reach for:

Building a new feature. From any directory: /aiforging:forge my-feature-name "brief description of what you want to build". It scaffolds a feature folder, pre-fills the spec, and walks you through refining scope and planning before any code is written. (Full form: /aiforging:new-feature.)

Reviewing code and catching a lesson. You're pair-programming with Claude, spot something less than ideal, and explain the fix. Claude's capture-pattern skill activates and offers to persist the lesson as a new pattern or anti-pattern file — one correction, one file, immediately available on every future Hammer pass. It'll ask whether the pattern applies to just this repo or all same-stack targets.

Hammer passes run automatically. At the end of each TDD cycle, hammer-refactor auto-triggers against the changed files — one fresh-context subagent per applicable pattern, parallel, isolated, no drift. You review each proposal and accept or reject. (You can also invoke it manually against any set of files if you want a targeted pass outside the normal flow.)

Auditing a codebase you just inherited. Onboard it with /aiforging:setup, and the architecture analyzer produces a scored assessment with prioritized findings. A structured starting point instead of grepping around.

Planning work that spans repos. From your forge workspace, /aiforging:forge tax-inclusive-pricing creates a single feature folder with one spec covering both backend and frontend targets — slices tagged per-repo, [gate: contract] on the API boundary. One plan, no drift between repos.

Who this is for

Software crafters with established codebases who already feel the pain of AI-generated sprawl and are ready to adopt an opinionated workflow. Teams whose backends are built around a Data Mapper ORM (Doctrine, Hibernate, Entity Framework Core, TypeORM, MikroORM) will get the most value out of the box. Teams on Active Record stacks (Eloquent, Rails) can still adopt the framework with caveats documented in the conventions.

AI Forging is not for greenfield projects in v0.2. A separate /aiforging:new-project command may come later. It is also not descriptive — it is prescriptive, and it will tell you to refactor things. That's the point.

The three-layer model

AI Forging deliberately separates three locations with different lifecycles: