dev

dev — Claude Code workflow harness plugin

A structured agent workflow harness for Claude Code. Provides skills and agents for investigating tasks, planning, reviewing, implementing, and learning from completed work.

What it does

The plugin provides a full development workflow:

1. Ideate (optional) — turn a vague problem or half-formed idea into a concrete next step
2. Investigate — pick from the project plan, or scope a user-provided task
3. Size — choose a workflow tier (minimal, light, full) appropriate to the task
4. Plan — produce planning documents matched to the tier
5. Review — spawn reviewer agents in parallel, loop until clean (default in full; opt-in in light)
6. Implement — execute the plan with mode-appropriate review checkpoints
7. Learn — process issues into documentation improvements (light, full)

Plus supporting skills:

• greenfield — set up a new project from scratch with standard conventions
• audit — check an existing project for harness readiness, scaffold gaps
• prepare-pr — clean up git history and create a pull request

Skills

Skill	Purpose
/dev:next [task]	Run the workflow for the next plan item, or for a task description given as argument
/dev:ideate	Turn a vague problem or half-formed idea into a concrete next step
/dev:investigate	Propose the next task to work on (from plan or from argument)
/dev:plan	Create planning documents (light: plan only; full: PRD + plan)
/dev:review-plan	Review a plan with all discovered reviewers (full mode by default; opt-in from light mode)
/dev:implement	Execute an approved plan with mode-appropriate review checkpoints
/dev:review-impl	Review implementation code with all discovered reviewers
/dev:learn	Process the issues journal into doc improvements
/dev:prepare-pr	Clean up history, push, and create a PR
/dev:greenfield	Bootstrap a new project with standard conventions
/dev:audit	Check harness readiness and scaffold missing pieces

Agents

Research (read-only, spawned during planning)

• codebase-researcher — explore current code state
• docs-and-learnings-researcher — find relevant documentation
• framework-researcher — fetch framework/library docs
• topic-researcher — targeted web research

Review (read-only, spawned at review checkpoints)

Always applicable:

• correctness-reviewer, simplicity-reviewer, security-reviewer, consistency-reviewer, architecture-reviewer, docs-reviewer, spec-compliance-reviewer

Stack-specific:

• rust-reviewer, frontend-reviewer

The plugin's reviewers run by default in any consuming project (no copy needed). A project can add its own reviewers under .claude/agents/review/ — same-name files override the plugin reviewer; differently-named files are additive. To drop a plugin reviewer entirely without replacing it, list its name under a ## Disabled reviewers section in the project's CLAUDE.md. Each reviewer reads project-specific criteria from .claude/conventions/ files.

Learning

• doc-improver — triage implementation issues into documentation fixes

Worker

• implement-worker — stateless per-step executor: writes tests + code, runs tests, returns a ~200-word report, context discarded after each step

/dev:next invokes downstream skills (/dev:plan, /dev:review-plan, /dev:implement, /dev:review-impl) via the Skill tool. Skill runs the called skill body in the caller's context (no subagent boundary), so the chain stays in the orchestrator's window. Each skill body then Agent-spawns its workers and reviewers — a single depth-1 spawn from the orchestrator, which is the depth Claude Code currently supports reliably.

Installation

# Add the marketplace
claude plugin marketplace add BalduinLandolt/dev-claude-plugin

# Install the plugin
claude plugin install dev@dev-claude-plugins

Or add to your project's .claude/settings.json:

{
  "enabledPlugins": {
    "dev@dev-claude-plugins": true
  },
  "extraKnownMarketplaces": {
    "dev-claude-plugins": {
      "source": {
        "source": "github",
        "repo": "BalduinLandolt/dev-claude-plugin"
      }
    }
  }
}

Project setup

The plugin expects consuming projects to provide: