mnemonic-rpi-workflow

分类: 工具与效率 | 上传者: danielmarbach | 下载: 0 | 版本: v1.0（最新）
在执行多步骤研究-计划-实施-审核工作流程时使用，特别是当您需要明确的移交、一份当前计划注释和一致的角色/关系约定时。
更新日志: Source: GitHub https://github.com/danielmarbach/mnemonic
目录结构

当前层级: tree/main/skills/mnemonic-rpi-workflow/
📄 SKILL.md 5.2 KB
SKILL.md

---
name: mnemonic-rpi-workflow
description: Use when executing multi-step research-plan-implement-review workflows, especially when you need explicit handoffs, one current plan note, and consistent role/relationship conventions.
---

# Skill: mnemonic-rpi-workflow

# RPIR Workflow: Research -> Plan -> Implement -> Review

Use this skill when work needs structured workflow artifacts.

## When to Use

- Multi-step feature or bugfix work where you need research, plan, execution, and review artifacts.
- Work delegated to subagents that needs explicit handoff and continuity.
- Tasks where plan drift risk is high and you want one current plan note.

## Core Principle

mnemonic is the canonical store for workflow artifacts, not the workflow runtime.

## Stage Checklists

### 1. Research

- Create or update one request root note with `role: context`, `lifecycle: temporary`, `tags: ["workflow", "request"]`.
- Before creating new notes, call `recall` for related prior work.
- Create research notes with `role: research`, `lifecycle: temporary`.
- Distill a short research summary when findings are scattered.
- Link research notes to request root. Prefer `derives-from` when lineage is explicit; otherwise use `related-to`.

### 2. Plan

- Create or update one current plan note with `role: plan`, `lifecycle: temporary`.
- Link plan note to request root and key research notes. Prefer `derives-from` for lineage and `follows` for sequence; fallback to `related-to` when direction is unclear.
- Keep the plan concise, executable, and scoped to the current request.
- Prefer one current plan per request; update or supersede as needed.
- If the plan changes materially, update the plan note before continuing implementation.

Material plan changes include architecture direction, file/module scope, ordering/dependencies, validation strategy, and key assumptions.

### 3. Implement

- Create apply/task notes as `lifecycle: temporary`, tagged with `apply`.
- Use `role: plan` for intended executable steps.
- Use `role: context` for observations, checkpoints, and execution notes.
- Link apply/task notes to the plan note. Prefer `follows` for ordered execution steps.
- For non-trivial work, dispatch a subagent with narrow scope and explicit handoff context.

### 4. Review

- Create review notes with `role: review`, `lifecycle: temporary`.
- Link review notes to apply/task notes or plan. Use `derives-from` when review conclusions derive from specific artifacts.
- Record outcomes: continue, block, or update plan.
- If review causes a material plan change, update plan note first.

### 5. Consolidate

- Create permanent decision note(s) for resolved approaches.
- Create permanent summary note(s) for outcomes and verification.
- Promote reusable patterns into permanent reference notes.
- Let pure temporary scaffolding expire when no longer useful.

## Subagent Handoff Template

Use this template for subagent delegation:

```text
Request note:
- <note-id/title>

Current plan note (or relevant slice):
- <note-id/title>

Relevant research notes:
- <note-id/title>
- <note-id/title>

Durable recalled context:
- <note-id/title>
- <note-id/title>

Task scope:
- Files/modules: <paths>
- Goal: <what to change>
- Validation: <tests/checks>

Subagent must return:
1) Updated apply note content
2) Optional review note content
3) Recommendation: continue | block | update plan
```

## Note Conventions

- Request root: one per workflow, `role: context`, `lifecycle: temporary`, `tags: workflow/request`.
- Apply/task split: no new role; use existing `plan` and `context` roles plus `apply` tag.
- Relationship density: keep sparse; link only to immediate upstream artifacts.
- Plan currency: one current plan note per request.

Relationship preference order:

1. `derives-from` for explicit lineage.
2. `follows` for explicit sequence.
3. `related-to` only when direction is unknown or intentionally generic.

Canonical graph:

```text
request root (context, temporary)
  -> research (temporary)
  -> plan (temporary)
  -> apply/task (temporary; plan for steps, context for observations)
  -> review (temporary)
  -> outcome/decision/summary (permanent)
```

## Commit Discipline

Use three commit classes through the workflow:

1. Memory commit: research/plan/review artifacts.
2. Work commit: code/test/docs implementation.
3. Memory commit: consolidation and promotion of durable knowledge.

If plan changes materially during implementation, do memory update first, then continue work.

## Examples

### Example 1: Research-heavy bug investigation

- Create request root + two research notes.
- Link research to request.
- Write one plan note with validation steps.
- Execute with apply notes and one review note.
- Consolidate into permanent decision + summary notes.

### Example 2: Multi-file refactor with subagent

- Create request root and concise plan note.
- Hand subagent request/plan/research plus narrow file list.
- Receive updated apply note + recommendation.
- Add review note, update plan if material drift occurred.
- Promote reusable patterns into reference note.

### Example 3: Small task with no iteration

- Create/update request root and one plan note.
- Implement directly with one apply note.
- Review note says continue and close.
- Consolidate to one permanent summary note.
登录后下载/点赞/收藏 ❤ 20 | ★ 0
mnemonic-rpi-workflow

目录结构

SKILL.md

举报内容

提示
