Claude Code Plugins

Community-maintained marketplace

Feedback

rpi-workflow

@matt-riley/mjrwtf
0
0

Apply a Research → Plan → Implement → Validate workflow with Copilot CLI, including context budgeting and persistent artifacts via /share.

Install Skill

1Download skill
2Enable skills in Claude

Open claude.ai/settings/capabilities and find the "Skills" section

3Upload to Claude

Click "Upload skill" and select the downloaded ZIP file

Note: Please verify skill by going through its instructions before using it.

SKILL.md

name rpi-workflow
description Apply a Research → Plan → Implement → Validate workflow with Copilot CLI, including context budgeting and persistent artifacts via /share.
license MIT
compatibility github-copilot-cli (interactive); uses /context, /clear, /share, @file mentions.
metadata [object Object]

Research → Plan → Implement → Validate (RPI-V)

Use this workflow for any non-trivial change (features, bugfixes, refactors) to keep Copilot CLI’s context focused and your changes predictable.

Core rules

  • Research first. No code changes during research.
  • Keep context under ~60%. Use /context frequently.
  • Clear between phases. Use /clear after you’ve saved the output for the next phase.
  • Persist artifacts to files so you can resume later (and so future work can reference them).
  • Implement one phase at a time, then run checks.

Recommended artifact locations

Create (locally) a thoughts/ directory (optionally symlinked to a global shared folder) and keep:

  • thoughts/shared/research/NNN_<topic>.md (or YYYY-MM-DD_<topic>.md)
  • thoughts/shared/plans/NNN_<topic>.md (or YYYY-MM-DD_<topic>.md)
  • thoughts/shared/validation/NNN_<topic>.md (or YYYY-MM-DD_<topic>.md)
  • thoughts/shared/sessions/NNN_<topic>.md (or YYYY-MM-DD_<topic>.md) (checkpoints / handoffs)

Optional (HumanLayer-style): keep a thoughts/searchable/ directory with hard links for fast searching.

This repo’s .gitignore excludes thoughts/ by default (treat it as local working memory).

How to run the workflow in Copilot CLI

Phase 1 — Research

Goal: locate relevant code, patterns, tests, and constraints.

  1. Ask focused questions and include only essential files with @path/to/file.
  2. Prefer targeted searches over dumping lots of files into context.
  3. Save the research notes:
    • /share file thoughts/shared/research/<id>_<topic>.md
  4. Check context (/context) and then /clear.

Suggested prompt template: see references/prompts.md.

Phase 2 — Plan

Goal: produce a phased plan with success criteria and commands.

  1. Start from a clean context.
  2. Provide the research artifact:
    • @thoughts/shared/research/<id>_<topic>.md
  3. Iterate until the plan is executable and phase-based.
  4. Save:
    • /share file thoughts/shared/plans/<id>_<topic>.md
  5. /context then /clear.

Phase 3 — Implement

Goal: execute the plan, one phase at a time.

  1. Start from a clean context.
  2. Provide the plan artifact and specify the phase:
    • @thoughts/shared/plans/<id>_<topic>.md
  3. Implement Phase 1 only, then run checks.
  4. If context approaches ~60%, save a checkpoint (session summary) to thoughts/shared/sessions/ via /share file thoughts/shared/sessions/<id>_<topic>.md, then /clear.

Phase 4 — Validate

Goal: prove the implementation matches the plan and repo conventions.

  1. Start from a clean context.
  2. Provide the plan artifact and what phases were implemented.
  3. Run repo verification commands (typically make test, go test ./..., make check as applicable).
  4. Record:
    • what matches the plan
    • deviations (intentional/unintentional)
    • remaining TODOs
  5. Save:
    • /share file thoughts/shared/validation/<id>_<topic>.md

For team sharing, you can also /share gist and paste the resulting link into issues/PRs.

Repo-specific verification defaults (mjr.wtf)

  • If SQL/templ changed: make generate
  • Tests: make test
  • Build sanity: go build ./...

If lint fails but tests/build pass, follow repo guidance (known lint false positives).