prove

prove

Plan, Research, Orchestrate, Validate, Execute — a Claude Code plugin that provides a complete plan-to-implementation lifecycle for any tech stack.

What It Does

Takes you from idea to merged code through a structured pipeline:

/prove:brainstorm  →  /prove:task-planner  →  /prove:plan-step  →  /prove:orchestrator  →  /prove:review  →  /prove:comprehend  →  /prove:cleanup
      │                        │                         │                        │                        │                        │                        │
.prove/decisions/       .prove/TASK_PLAN.md     .prove/plans/plan_X/    .prove/reports/         .prove/reviews/         .prove/learning/         .prove/archive/

1. Brainstorm — Explore options, weigh trade-offs, record decisions
2. Task Planner — Discover requirements via questionnaires, create incremental plans
3. Plan Step — Deep-dive into individual steps: requirements, design decisions, test strategy
4. Orchestrator — Autonomous execution with validation gates and git snapshots
5. Review — Generate an Agent Change Brief for structured code review
6. Comprehend — Socratic quiz on agent-generated diffs to build code comprehension
7. Cleanup — Archive artifacts, remove working files, delete branches

Key Features

• Auto-scaling orchestrator: Small tasks (≤3 steps) run sequentially. Larger tasks use parallel git worktrees with mandatory architect review. Three execution modes: /prove:orchestrator, /prove:autopilot, and /prove:full-auto.
• Structured code review: Intent-manifest-driven code review. Agents declare why they made each change via Claude Code hooks. Changes are grouped by intent for structured review in a browser-based UI. Run /prove:review to start.
• Code quality steward: Deep line-by-line audits (/prove:steward), iterative audit-fix loops (/prove:steward:auto-steward), and lightweight session-scoped reviews (/prove:steward:steward-review). See docs/code-quality.md.
• Stack-agnostic validation: Auto-detects your project type (Go, Rust, Python, Node, Godot, Makefile) and runs appropriate build/lint/test checks. Supports LLM-based prompt validators for higher-level checks. Configure with .claude/.prove.json or let auto-detection handle it.
• Session management: Create handoff prompts (/prove:handoff) to preserve context across Claude Code sessions. Resume with /prove:pickup in a fresh session.
• Documentation generation: Human-readable docs (/prove:docs), LLM-optimized agent docs (/prove:docs:agentic-docs), or both at once (/prove:docs:auto-docs).
• Git-based rollback: Every step is committed individually. Revert any step, reset to any point.

Installation

claude plugin marketplace add mjmorales/claude-prove
claude plugin install prove@prove

If Claude Code is already running, restart it for the plugin to take effect.

Quick Start

# Initialize validation config for your project
/prove:init

# Start brainstorming a feature
/prove:brainstorm

# Plan the implementation
/prove:task-planner

# Deep-dive into a specific step
/prove:plan-step 1.2.3

# Execute autonomously
/prove:orchestrator

# Review agent-generated code
/prove:review

# Quiz yourself on what the agent wrote
/prove:comprehend

# Clean up when done
/prove:cleanup my-feature

Commands Reference

Planning & Discovery

Command	Description
/prove:brainstorm	Interactive brainstorming — explore options, record decisions to .prove/decisions/
/prove:task-planner	Guided requirements gathering and incremental plan creation
/prove:plan-step [step]	Deep-dive into a specific plan step: requirements, design, test strategy
/prove:spec [topic]	Author formal specifications following RFC/IETF conventions

Execution

Command	Description
/prove:orchestrator	Execute a task plan with validation gates. Auto-scales between simple and full mode
/prove:autopilot [plan]	Run the orchestrator hands-off on a specific plan
/prove:full-auto	End-to-end: requirements → plan → parallel execution → merge
/prove:prep-permissions	Pre-configure Claude Code permissions for smooth orchestrator runs

Code Review

Command	Description
/prove:review	Assemble intent manifests and launch the review UI
/prove:review:resolve	Show approval summary — accepted groups and merge readiness
/prove:review:fix	Generate fix prompts from rejected review groups
/prove:review:discuss	Surface groups needing discussion from review

Code Quality