RPI Workflow — Research → Plan → Implement

A structured alternative to BMAD. Three explicit phases with a GO/NO-GO gate between them. Each phase has its own subagents and produces a named artefact. Most useful for medium-to-large changes (≥1 day of work, ≥10 files) where ad-hoc planning loses the thread.

When to Use

Change size	Use
< 1 hour, < 5 files	Direct edit. Skip RPI.
1 hour – 1 day, single concern	Lightweight: plan in chat, execute.
≥ 1 day, multi-file, multi-concern	RPI
≥ 1 sprint, cross-team, regulatory	RPI + BMAD personas

The Three Phases

Phase 1 — Research (read-only)

Goal: understand the problem space and produce a GO/NO-GO recommendation. No code changes.

Subagents:

research-explorer — codebase reconnaissance (greps, file maps)
research-domain — pulls external docs / RFCs / vendor SDKs (Context7, web)
research-risk — identifies blockers, hidden coupling, regression vectors

Output: tasks/<slug>/research.md with sections Findings, Unknowns, Risks, GO/NO-GO.

Slash command: /rpi:research <slug>. Stop here if NO-GO. Do not let momentum carry you into a phase you should not be in.

Phase 2 — Plan (read-only)

Trigger: research returns GO. Three planning subagents run in parallel:

plan-pm → tasks/<slug>/pm.md — user-facing scope, success criteria, rollback story
plan-ux → tasks/<slug>/ux.md — interaction surfaces, copy, error states
plan-eng → tasks/<slug>/eng.md — module touch list, migration order, test strategy

Then plan-synth merges them into tasks/<slug>/PLAN.md and lists open questions. Loop back to PM/UX/eng if the synthesis surfaces a contradiction.

Slash command: /rpi:plan <slug>. The merged PLAN.md is the contract for phase 3.

Phase 3 — Implement

Trigger: PLAN.md approved. One executor agent at a time, code-reviewer agent runs after each chunk.

executor (sonnet) — works through the eng.md module list in order
code-reviewer — independent pass after each chunk, blocks merge on rated issues
verifier — runs the test plan from eng.md and posts pass/fail evidence

Slash command: /rpi:implement <slug>. Stop and re-plan if the executor hits a blocker not in research/plan — do not improvise.

Artefact Layout

tasks/
  <slug>/
    research.md      # phase 1
    pm.md ux.md eng.md   # phase 2 inputs
    PLAN.md          # phase 2 output (the contract)
    review-<chunk>.md  # phase 3, per chunk
    verify.md        # phase 3 final evidence

Commit each phase artefact before starting the next. Future sessions can pick up mid-flow because the contract is on disk.

RPI vs BMAD

	RPI	BMAD
Shape	Linear, gated	Persona-based, parallel
Best for	Medium changes with clear scope	Large changes needing role separation
Output	Single PLAN.md contract	Multiple persona artefacts
Stop conditions	NO-GO at gate	Persona consensus

They compose: RPI for the structure, BMAD personas as the agent identities inside each RPI subagent.

Pitfalls

Skipping the gate. If research returns NO-GO, do not start planning anyway because “we already paid the research cost”. Sunk-cost trap.
One mega-plan. If PLAN.md is more than ~200 lines, split the work — that’s a sign the change should be two RPI runs.
Executor without reviewer. Code-reviewer is not optional. Self-review in the same context is invalid.
Memory leak between phases. Phase 3 must not load research.md beyond what eng.md cites — context bloat kills implementation accuracy.