rpikit

A plugin implementing the Research-Plan-Implement (RPI) framework for disciplined software engineering.

Philosophy

Understand before acting. Plan before coding. Implement with discipline.

This plugin enforces a structured workflow that prevents premature implementation and ensures human oversight at
critical decision points.

Installation

Step 1: Add the marketplace

/plugin marketplace add bostonaholic/rpikit

Step 2: Install the plugin

/plugin install rpikit

Quick Start

Workflow

Full Pipeline

flowchart LR
    rpi["/rpikit:research-plan-implement"] -->|automated| research[Research]
    research -->|approval| plan[Plan]
    plan -->|approval| implement[Implement]
    implement --> done((done))

The /rpikit:research-plan-implement skill runs the full pipeline in a single session using parallel subagents, with
approval gates between phases. Implementation runs in an isolated worktree to protect the main branch.

Individual Skills

flowchart LR
    brainstorm["/rpikit:brainstorming"] -.->|optional| research["/rpikit:researching-codebase"]
    research -->|approval| plan["/rpikit:writing-plans"]
    plan --> implement["/rpikit:implementing-plans"]
    plan -.->|optional| decision["/rpikit:documenting-decisions"]
    implement -->|approval| done((done))

For more control, run each phase as a separate skill. Each phase produces artifacts in docs/plans/ and requires human
approval before transitioning to the next phase.

Brainstorming vs. Research

Both commands start by asking clarifying questions before acting. The key difference is their purpose:

Use Brainstorming when:

• Requirements are vague: "Add user auth" → What kind? OAuth? JWT? Sessions?
• Multiple approaches exist: "Improve performance" → Which areas? What trade-offs?
• Design decisions needed before you can research

Use Research when:

• You know what to build but need to understand the codebase
• "Where is authentication implemented?" → Finds files, traces flow
• "How does the existing caching work?" → Documents patterns

Common flow: Brainstorm first (if unclear) → Research → Plan → Implement

Output Structure

docs/plans/
├── YYYY-MM-DD-<topic>-research.md   # Research findings with file:line references
└── YYYY-MM-DD-<topic>-plan.md       # Implementation plan with tasks and criteria

docs/decisions/
└── NNNN-decision-title.md           # Architecture Decision Records

Usage Examples

Full Pipeline (Recommended)

Run the entire research-plan-implement workflow in a single session:

/rpikit:research-plan-implement Add OAuth login with Google and GitHub providers

This spawns parallel research subagents, synthesizes findings, creates a plan, and implements it — with approval gates
between each phase.

Step-by-Step Workflow

For more control, run each phase separately:

/rpikit:researching-codebase I want to add OAuth login - what auth patterns exist?

Review the research output in docs/plans/, then create a plan from it:

/rpikit:writing-plans @docs/plans/2025-01-07-oauth-login-research.md

Review and approve the plan, then implement from it:

/rpikit:implementing-plans @docs/plans/2025-01-07-oauth-login-plan.md

Ad-hoc Code Review

Review current changes for quality issues:

/rpikit:reviewing-code

Review for security vulnerabilities:

/rpikit:security-review

Recording Decisions

After planning or design work, record the decision as an ADR:

/rpikit:documenting-decisions @docs/plans/2025-01-07-oauth-login-design.md

Stakes-Based Planning

The framework adapts to change complexity:

• Low stakes (docs, formatting): Minimal planning, quick execution
• Medium stakes (new features, refactors): Full RPI workflow
• High stakes (architecture, security): Thorough research and detailed planning

Skills

The plugin includes methodology skills that guide disciplined development:

Core RPI Workflow

• research-plan-implement - End-to-end pipeline orchestrating all RPI phases with parallel subagents
• researching-codebase - Thorough codebase research through interrogation
• synthesizing-research - Consolidate parallel research findings into a unified report
• writing-plans - Granular, verifiable implementation plans
• implementing-plans - Disciplined execution with checkpoint verification
• reviewing-code - Quality review using Conventional Comments
• security-review - Security-focused review for vulnerabilities

Development Discipline

• test-driven-development - RED-GREEN-REFACTOR cycle enforcement
• systematic-debugging - Root cause investigation before fixes
• verification-before-completion - Evidence before claims

Workflow Support

• brainstorming - Collaborative design before research/planning
• documenting-decisions - Record architectural decisions as ADRs
• finishing-work - Structured completion (merge, PR, cleanup)
• receiving-code-review - Verification-first response to feedback

Advanced Patterns

• git-worktrees - Isolated workspaces for parallel work
• parallel-agents - Concurrent dispatch for independent tasks

Architecture

For a detailed overview of how skills, agents, and hooks connect, see Architecture.

Inspired By

• No Vibes Allowed - Dex Horthy's talk on context engineering and the RPI
  workflow for solving hard problems in complex codebases
• superpowers - Composable skills
• BMAD Method - Scale-adaptive planning
• SuperClaude Framework - Behavioral modes and deep research
• RPI Framework - RPI structure
• HumanLayer - Human-in-the-loop approval patterns

License

MIT