Now liveThe Skillselion MCP - thousands of ranked skills, loaded into your agent mid-task. No install.Get it →
darven avatar

P8 Engine

  • 13 repo stars
  • Updated July 14, 2026
  • darven-cs/p8-engine

Apply an engineering methodology with an AI memory system: core rules, quality self-checks, emergency protocol, and lightweight memory injection.

About

p8-engine bundles an engineering methodology with a lightweight AI memory system across two skills: core rules with quality self-checks and an emergency protocol, plus memory injection using templates, hooks, and atomic commits. Developers use it to keep an agent's engineering practice consistent and to give it durable project memory while developing.

  • Engineering methodology rules
  • AI memory injection
  • Quality self-check
  • Emergency protocol

P8 Engine by the numbers

  • Data as of Jul 15, 2026 (Skillselion catalog sync)
/plugin marketplace add darven-cs/p8-engine
/plugin install p8-engine@p8-engine

Add your badge

Show developers this skill is listed on Skillselion. Paste this into your README.

Listed on Skillselion
repo stars13
Last updatedJuly 14, 2026
Repositorydarven-cs/p8-engine

What it does

Apply an engineering methodology with an AI memory system: core rules, quality self-checks, emergency protocol, and lightweight memory injection.

README.md

P8 Engine

工程方法论 + AI 记忆系统 —— 一个 plugin,两个 skills,全面提升 AI 编码质量。

中文 | English


中文

这是什么

P8 Engine 是一个 Claude Code Plugin,包含两个 skills:

Skill 用途 触发方式
p8-engine 工程方法论 — 铁律、链路预判、行动后自检、全链路审视、能动性鞭策 /p8-engine
inject-memory 给任意项目注入轻量 AI 记忆系统 /inject-memory

安装

# 添加 marketplace
claude plugin marketplace add darven-cs/p8-engine

# 安装 plugin(一次安装,两个 skills 都可用)
claude plugin install p8-engine@p8-engine

# 激活(启动新会话后)
/p8-engine          # 激活工程方法论
/inject-memory      # 注入记忆系统到目标项目

p8-engine:工程方法论

不是 prompt wrapper。是行为协议,让 AI 像 P8 工程师一样做事。

核心机制
会话启动协议    风格检测 → 任务校准 → 链路预判 → 确认激活
核心铁律        穷尽方案 · 先查后问 · 主动延伸 · 全链路排查 · 精准修改 · 检查点意识
行动后自检      每个动作后立即验证 + 战果记录 + 能动性鞭策
质量自检        6 项通用 + 5 项代码 + 交付前全链路审视
应急流程        诊断卡壳 → 拉高视角 → 最小行动恢复 → 穷尽后结构化退出
模块开发        Phase 0-4:读 → 设计契约 → 自检 → 实现 → 验证
它治什么病
LLM 编码常见问题 P8 Engine 怎么治
空口说"已完成",不跑验证 战果记录:每个完成都必须有 build/test/curl 输出
修完 bug 就停,不检查同类 铁律三 + 行动后自检:同 root cause 范围内延伸排查
信息不足就问用户 铁律二:先用工具自查,只问真正需要确认的
只修当前报错,不看全链路 铁律四 + 链路预判 + 全链路审视:自底向上画依赖图
乱猜假设直接跑 铁律:有歧义先澄清,不代入假设
把 100 行写成 1000 行 质量自检:简洁性检查,不加未被要求的抽象
顺手改了无关代码 铁律五:精准修改,不做附带"改进"
代码风格和项目格格不入 风格检测:开工前强制读现有代码
卡壳反复微调同一方案 应急流程:诊断 → 最小行动 → 渐进恢复
修完就停,不验证不延伸 能动性鞭策:"端到端在哪?证据呢?"
交付质量参差不齐 质量罗盘 + 全链路审视:交付前强制逐跳验证

inject-memory:AI 记忆系统

解决一个核心问题:AI 每次新会话都失忆。上一次会话踩过的坑、做的设计决策、模块间的依赖关系——全忘了,下次从头踩。

一键注入
cd /path/to/your-project
/inject-memory

注入后项目新增:

your-project/
├── memory/
│   ├── _index.md                    # 总索引(每次会话自动加载)
│   ├── _tutorial.md                 # 使用教程(可复制到新项目)
│   ├── modules/                     # 模块记忆
│   │   ├── overview.template.md     # 模块概览模板
│   │   ├── design.template.md       # 设计决策模板
│   │   ├── integration.template.md  # 接入方式模板
│   │   └── known_issues.template.md # 已知问题模板
│   ├── bugs/                        # 踩坑记录
│   │   └── bug.template.md          # 踩坑模板
│   └── progress/                    # 进度追踪
│       └── current-status.template.md
├── .claude/
│   ├── hooks/
│   │   └── sync-memory-index.mjs    # PostToolUse:改记忆后自动同步索引日期
│   └── settings.json                # hook 注册
└── CLAUDE.md                        # 追加了记忆使用规范
它怎么工作

每次新会话启动时:Claude 自动加载 memory/_index.md,感知项目全貌——有哪些模块、踩过哪些坑、当前进度是什么。

每次改记忆后:PostToolUse hook 自动把 _index.md 对应条目的"最后更新"日期同步为今天(一次可更新所有匹配行)。

记忆不会丢:结构化模板 + 索引 + hook 三件套,跨会话持久化。

什么该记,什么不该记
该记(项目记忆) 不该记
模块架构、接口、数据流 代码本身的细节(读代码就行)
设计决策和为什么这样选 临时调试日志
踩过的坑(跨模块、代码看不出原因的) 一次性脚本的用途
模块间依赖关系和集成点 API 文档(外部链接即可)
当前进度和未完成事项 已完成且不再相关的任务
环境特定的配置约束 通用编程知识

判断标准:下次会话读到这条记忆,能不能避免重复犯错或重复调研? 能 → 记。不能 → 不记。

记忆目录结构

模块记忆memory/modules/)—— 一个模块一份记忆,大模块用子目录:

modules/
├── auth/                          # 大模块 → 子目录
│   ├── overview.md                # 有什么类/接口/数据
│   ├── design.md                  # 为什么这样做
│   ├── integration.md             # 和其他模块怎么对接
│   └── known_issues.md            # 未修的坑
├── quick-utils.md                 # 小模块 → 单文件
└── ...

踩坑记录memory/bugs/)—— 跨模块的通用坑:

bugs/
├── redis_cache-invalidation.md    # {技术栈}_{现象}.md
├── nginx_proxy-websocket.md
└── ...

进度追踪memory/progress/)—— 会话间的状态传递:

progress/
├── current-status.md              # 当前在做什么
└── session-2025-01-15.md          # 历次会话记录
和 Claude Code 自带 memory 的区别
Claude Code auto memory inject-memory
存储位置 ~/.claude/ 用户级 项目 memory/ 目录
谁能读 只有你自己 所有项目协作者(git 管理)
结构 扁平 markdown 索引 + 模板 + 分类目录
自动化 PostToolUse hook(自动同步索引日期)
适用场景 个人偏好、通用习惯 项目架构、设计决策、踩坑记录

两者互补:auto memory 记你是什么样的开发者,inject-memory 记这个项目是什么样的项目。


两个 skills 怎么配合

/inject-memory          # 第一步:给项目注入记忆系统(一次性)
/p8-engine              # 第二步:激活工程方法论(每次会话)
                        # 之后正常开发

p8-engine 的行动后自检产出的战果,可以沉淀到 inject-memory 的 memory/bugs/memory/modules/ 中。

inject-memory 的记忆在下次会话被 p8-engine 的启动协议加载(步骤一.五:经验记忆加载),形成闭环:

会话 1:开发 → 行动后自检发现坑 → 写入 memory/bugs/
会话 2:启动 → 加载 memory/_index.md → 自动规避已知坑

配置 Claude Code

方式一:Plugin 安装(推荐)
# 1. 添加 marketplace
claude plugin marketplace add darven-cs/p8-engine

# 2. 安装
claude plugin install p8-engine@p8-engine

# 3. 使用(新会话中)
/p8-engine
/inject-memory
方式二:手动配置

如果 plugin 系统不可用,手动将 skills 复制到项目:

# 1. 克隆仓库
git clone https://github.com/darven-cs/p8-engine.git /tmp/p8-engine

# 2. 复制 skills 到项目
mkdir -p .claude/skills
cp -r /tmp/p8-engine/plugins/p8-engine-plugin/skills/p8-engine .claude/skills/
cp -r /tmp/p8-engine/plugins/p8-engine-plugin/skills/inject-memory .claude/skills/

# 3. 复制 commands
mkdir -p .claude/commands
cp /tmp/p8-engine/plugins/p8-engine-plugin/commands/*.md .claude/commands/
安装后验证
# 检查 skills 是否就位
ls .claude/skills/p8-engine/SKILL.md
ls .claude/skills/inject-memory/SKILL.md

# 检查 commands 是否就位
ls .claude/commands/p8-engine.md
ls .claude/commands/inject-memory.md

# 在 Claude Code 中测试
/p8-engine       # 应输出 "P8 模式已激活"
/inject-memory   # 应开始执行注入流程

使用教程

完整教程

  1. 注入 p8-engine,激活后执行任务 案例1

  2. 风格缓存命中,减少扫描次数 首次记忆加载

  3. Bug 修复时全链路排查,避免 AI 乱跑 案例2 案例2

  4. 复杂问题先扫描项目再调研 案例3

项目结构

p8-engine/
├── .claude-plugin/
│   └── marketplace.json              # marketplace 注册
├── plugins/
│   └── p8-engine-plugin/             # 唯一的 plugin
│       ├── .claude-plugin/
│       │   └── plugin.json           # plugin 清单
│       ├── skills/
│       │   ├── p8-engine/            # 工程方法论 skill
│       │   │   ├── SKILL.md
│       │   │   └── references/
│       │   └── inject-memory/        # 记忆系统注入 skill
│       │       ├── SKILL.md
│       │       ├── hooks/
│       │       └── templates/
│       └── commands/
│           ├── p8-engine.md
│           └── inject-memory.md
├── docs/
├── README.md
└── LICENSE

English

What is this

P8 Engine is a Claude Code Plugin with two skills:

Skill Purpose Trigger
p8-engine Engineering methodology — iron rules, chain preview, post-action review, full-chain audit, agency enforcement /p8-engine
inject-memory Inject lightweight AI memory system into any project /inject-memory

Installation

# Add marketplace
claude plugin marketplace add darven-cs/p8-engine

# Install plugin (one install, both skills available)
claude plugin install p8-engine@p8-engine

# Activate (after starting new session)
/p8-engine          # Activate engineering methodology
/inject-memory      # Inject memory system into target project

p8-engine: Engineering Methodology

Not a prompt wrapper. A behavioral protocol that makes AI work like a senior engineer.

LLM coding problem How P8 Engine fixes it
Claims "done" without verification Battle records: every completion must have build/test/curl output
Fixes a bug and stops Iron Rule 3 + post-action review: sweep same root cause
Asks user when info is insufficient Iron Rule 2: investigate with tools first
Only fixes the immediate error Iron Rule 4 + chain preview + full-chain audit: bottom-up dependency chain
Makes wrong assumptions Iron Rule: surface ambiguity before coding
Inflates 100-line solutions to 1000 lines Quality check: simplicity, no unrequested abstractions
Makes collateral changes Iron Rule 5: surgical changes only
Code style clashes with project Style detection: mandatory code reading before writing
Stuck tweaking same approach Emergency protocol: diagnose → minimal action → recover
Stops after fix, no verification Agency enforcement: "Where's the end-to-end? Evidence?"

inject-memory: AI Memory System

Solves one core problem: AI forgets everything between sessions. Pitfalls, design decisions, module dependencies — all gone.

cd /path/to/your-project
/inject-memory

After injection:

  • memory/ directory with structured templates (modules/bugs/progress)
  • .claude/hooks/ with a PostToolUse hook (index sync)
  • CLAUDE.md appended with memory usage rules

Every new Claude session loads memory/_index.md automatically. The PostToolUse hook syncs index timestamps after memory updates.

What to remember
Remember (project memory) Don't remember
Module architecture, interfaces, data flow Code details (just read the code)
Design decisions and why Temporary debug logs
Cross-module pitfalls One-time script purposes
Module dependencies and integration points External API docs (link instead)
Current progress and TODOs Completed tasks
Environment-specific constraints General programming knowledge

Rule of thumb: Will this memory prevent duplicate mistakes or duplicate research next session? Yes → remember. No → don't.

How the two skills work together

/inject-memory          # Step 1: inject memory system (once)
/p8-engine              # Step 2: activate methodology (every session)
                        # Then work normally

p8-engine's post-action review produces battle records that flow into inject-memory's memory/bugs/ or memory/modules/. Next session, p8-engine's startup protocol loads these memories, closing the loop.

License

MIT

Related skills

This week in AI coding

Five minutes, every Monday - the tools, releases and tactics for developers.

unsubscribe anytime.