360 Secure SkillHub
360 安全技能中心

brag

面向个人贡献者的成果记录系统:通过 Context、Capture、Compiler 三类 agent,在年度工作区 brag-YYYY/ 中建立公司/角色上下文、按情境-行动-结果结构记录成果,并重新生成 brag 文档。强调不虚构、不夸大,保存前需用户确认,用于绩效评估、晋升与职业沟通。

gabepauli
热度 0.00
最近更新 2026-10-18 17:14:17
分类:知识依赖:读写文件 / 内部调用大模型应用场景:笔记整理 / 报告撰写 / 工作流 / 模板

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add gabepauli/brag-skill/brag
Metadata
name
brag
description
A brag document system for individual contributors. Helps anyone capture, classify, and compile their achievements into a living document for performance reviews, promotions, and career conversations. Use this skill when the user wants to log an achievement, set up their brag doc workspace, compile their brag document, or track their work over time. Trigger whenever the user mentions something they shipped, a project they worked on, something worth documenting, or asks about their accomplishments. Also trigger for phrases like "I did something today", "log this", "add to my brag doc", "compile my work", or "set up my workspace".

Brag

A system for individual contributors to capture achievements as they happen and compile them into a living brag document. Honest, specific, and portable across companies and roles.

The system has three agents. Each has a dedicated file with full instructions. Read the relevant agent file before acting.


Agent Directory

Agent File When to invoke
Context agents/context-agent.md Setup, new company, role change
Capture + Analyze agents/capture-agent.md Logging an achievement
Compiler agents/compiler-agent.md Generating the brag doc

Routing

Read the message and route to the right agent. Don't ask which agent they want — infer it.

Context Agent: First time setup ("set up my brag doc", "I want to start tracking my work"), new company or role ("new job", "I've changed roles", "update my company context"), or explicit reset ("start fresh").

Capture + Analyze Agent: Logging work ("log an achievement", "I want to brag about something"), describing work unprompted ("I just shipped X", "I ran sessions for Y"), or anything that sounds like something the user did recently.

Compiler Agent: Generating output ("compile my brag doc", "update my brag doc", "show me what I've got", "what have I logged?").

When ambiguous, default to Capture + Analyze. Someone describing their work is almost always trying to log it.

When no workspace exists: Before routing to any agent, if no brag-YYYY/ folder is found in the current directory, show this message and route to the Context Agent:

Brag helps you document your work as it happens, so you're never starting from scratch at review time.

I can see there are no context files created in your folder, so I'll ask a few questions first.


Workspace structure

The brag-YYYY/ folder lives in the current working directory. All agents read from and write to this folder. A new folder is created each year — previous years are never modified.

brag-2025/
    company-skill-matrix.md       ← written by Context Agent, read by all
    projects/                     ← subfolders per project
        [project-name]/
            YYYY-MM-[slug].md     ← one entry per file
    standalone-wins/
        YYYY-MM-[slug].md
    influence-and-collaboration/
        YYYY-MM-[slug].md
    learning-and-growth/
        YYYY-MM-[slug].md
    outside-work/
        YYYY-MM-[slug].md
    brag-doc.md                   ← compiled output, always regenerated

Core principles

These apply to every agent in the system.

Never fabricate. Every word in every entry comes from the user. If information is missing, ask for it or flag it — never invent it.

Never inflate. Qualitative outcomes are as valid as quantitative ones. "The doc became the go-to reference" is a complete result. No fake precision, no significance inflation.

User owns the writing. Agents propose, users approve. Nothing is saved without explicit confirmation.

Entries are the source of truth. The markdown files in the bucket folders are permanent records. The brag doc is always regenerated from them — never edited directly.

Keep it moving. The system should feel like a fast, useful conversation. Two rounds of back-and-forth to capture an entry is ideal. Four is the maximum.


Writing standards

All agents apply these when producing any text shown to the user or saved to disk. Based on the humanizer principles from blader/humanizer.

Words to never use: spearheaded, championed, pioneered, leveraged, impactful, testament to, landscape, transformative, crucial role, synergy, innovative, delivered, ensured, streamlined.

Words to use: specific verbs. Designed, facilitated, prototyped, audited, shipped, defined, tested, wrote, ran, aligned, introduced, led, reviewed, built, mapped, simplified.

Write in first person, past simple or present perfect. Short sentences. One idea per sentence. No em dash overuse. No bold inside prose. No bullet points inside SAR entries.


Reference files

Full agent instructions live in the agents/ directory. Always read the full agent file before running that agent.

  • agents/context-agent.md — workspace setup, skill matrix ingestion
  • agents/capture-agent.md — SAR capture, classification, project linking
  • agents/compiler-agent.md — brag doc generation
  • templates/entry-template.md — entry structure reference for parsing