当决策涉及所有者、客户和失败成本时，我正在培养我希望身边的代理技能。

我对那些看起来只是完成的输出几乎毫无兴趣。一份精心打磨的报告仍然可能掩盖一个猜测的工作流程、缺失的证据、不明确的权威，以及一个没人知道如何操作的系统。

这些技能让代理展示其工作：检查资料，提出未解决的问题，将判断与证据分开，测试失败案例，并生成与决策相关的产物。我可以阅读每条指令，不同意，并修改它们。

利用它们，挑战它们，并将它们适应你的工作。让做决定的人承担责任。

[!注] 感谢Matt Pocock，他的《真正工程师的技能》启发了本项目。Matt将软件工程实践转化为小型、可组合的技能;我正在将这一理念应用于角色特定工作，如客户发现、判断、交付和交接。

安装（30秒设置）

安装一个带有 skills CLI 的命名技能。本示例安装 fde-engagement：

安装程序会检测到你支持的代理，并询问该技能安装在哪里。

Skills To Do Work

I am building the Agent Skills I want beside me when a decision has an owner, a customer, and a cost of failure.

I have little use for output that only looks finished. A polished report can still hide a guessed workflow, missing evidence, unclear authority, and a system nobody knows how to operate.

These skills make the agent show its work: inspect source material, ask unresolved questions, keep judgment separate from evidence, test failure cases, and produce artifacts tied to a decision. I can read every instruction, disagree with it, and change it.

Use them, challenge them, and adapt them to your work. Keep the person making the decision accountable.

[!NOTE] With thanks to Matt Pocock, whose Skills for Real Engineers inspired this project. Matt turns software engineering practice into small, composable skills; I am applying that idea to role-specific work such as customer discovery, judgment, delivery, and handoff.

Installation (30-second setup)

Install a named skill with the open skills CLI. This example installs fde-engagement:

npx skills@latest add likamrat/skills --skill fde-engagement

The installer detects your supported agents and asks where to install the skill.

For a non-interactive GitHub Copilot project install:

npx skills@latest add likamrat/skills --skill fde-engagement --agent github-copilot -y

Then ask your agent:

Use fde-engagement to qualify this customer problem.
Start by asking for the smallest evidence package that unlocks the decision.

No additional project setup is required.

Installation commands live in this section. Skill READMEs link here instead of copying commands that can drift.

Why these skills exist

The agent starts building before the workflow is understood

A plausible architecture is easy to produce from a clean problem statement. Customer work rarely arrives that way. Operators disagree, systems record only part of the process, exceptions carry the risk, and the requested solution may solve the wrong problem.

fde-engagement starts with fit, outcome, baseline, operators, evidence, and constraints. It blocks architecture work until the current workflow and responsibility boundaries are explicit.

Polished prose fills gaps that a person should answer

An agent can make weak evidence sound settled. It can also invent the opinion, rationale, or confidence that gives a report its voice.