dev

ai-coding

Personal AI coding workspace — one repo for all AI dev tooling.

Includes a Claude Code plugin marketplace (dev orchestrator with parallel stack-aware agents) and Cursor tooling (code review orchestrator with parallel reviewer agents). Claude Code ignores .cursor/; Cursor ignores .claude-plugin/ and plugins/. Both coexist cleanly.

Claude Code Plugin — Dev Orchestrator

A streamlined 4-phase development workflow that parallelizes stack-aware agents and biases toward action.

Install

# From GitHub (anyone):
/plugin marketplace add DanielOchoa/ai-coding

# Local:
/plugin marketplace add <path-to-repo>/ai-coding

# Then install the plugin:
/plugin install dev@ai-coding

After marketplace install, commands become namespaced: /dev:run, /dev:scout, /dev:verify.

Commands

Command	Purpose
/dev:run [description]	Full 4-phase workflow: Context → Plan → Build → Verify + Ship
/dev:scout [area]	Quick codebase exploration with stack detection
/dev:verify [scope]	Post-implementation quality check + docs + commit

Agents

Agent	Model	Purpose
stack-detective	haiku	Fast project detection, reads CLAUDE.md, returns structured context
rails-specialist	sonnet	Rails 8 code tracing — MVC, ActiveRecord, Turbo/Stimulus
astro-specialist	sonnet	Astro 5.x code tracing — content collections, components, SSG
code-tracer	sonnet	Generic code path tracer — adapts to any stack as specialist fallback
tailwind-ui	sonnet	Tailwind CSS + frontend patterns, design system analysis
implementation-planner	opus	One decisive blueprint — files, order, tests, risks
bug-hunter	opus	Post-impl review — logic errors, edge cases (confidence >= 80)
convention-enforcer	haiku	CLAUDE.md compliance checking against the diff

How /dev:run Works

1. Context (~10s, parallel) — stack-detective + specialist + tailwind-ui (conditional)
2. Plan (~30s) — implementation-planner (opus) produces one decisive blueprint
3. Build (variable) — main session implements the plan, runs tests
4. Verify + Ship (~20s) — bug-hunter + convention-enforcer, then docs-updater + git-commit-author

Design Principles

• Generic agents, runtime context: Agents have stack-level expertise but zero business-specific knowledge. All project context comes from reading CLAUDE.md at runtime. Safe for a public repo.
• Right model for the job: Planner and bug-hunter run on opus (deep reasoning, inherits session effort — xhigh recommended). Tracers and UI analysis run on sonnet. Pure pattern-matching agents (stack-detective, convention-enforcer) run on haiku for speed and cost.
• Reference existing agents: docs-updater and git-commit-author are invoked from ~/.claude/agents/ — updates propagate automatically.
• 4 phases, not 7: No discovery phase, no 3-option architecture menus, no serial blocking questions.

Future Work

Potential improvements to explore once the current orchestrator-worker pattern has been stress-tested across more projects.