Announcing the AI Coding Agent Skills Guidebook for government teams

When someone joins your software team, they need context about how things work. AI coding tools need that context, too. Your agency’s coding standards, architecture decisions, security requirements, and business rules help these tools generate code that fits your team’s needs.

Skills and instruction files give teams a way to capture that knowledge and provide it to coding agents. Once that guidance is written down, you can review it, share it across teams, and update it as requirements change. Teams also need to test whether the guidance improves the code their tools produce.

We developed the Ad Hoc AI Coding Agent Skills Guidebook to help government agencies put those practices in place. It’s designed for teams developing their first skills as well as organizations coordinating their use across multiple teams.

Making agency knowledge available to coding agents

Tools like GitHub Copilot, Claude Code, and Cursor can use instructions that describe how a team’s software needs to work. Depending on the tool, those instructions might live in an AGENTS.md or CLAUDE.md file, an on-demand skill, or a custom agent configuration.

Whatever form the instructions take, it’s important that some of that knowledge comes from outside the software team. Policy experts, records officers, security staff, and other subject matter experts all hold knowledge that should inform the code an agent generates. Our guidebook explains how to involve them in developing and reviewing that guidance.

Practical guidance for developing and managing skills

That collaboration is one of nine practices we offer in this guidebook for developing skills, sharing them, evaluating their effectiveness, and keeping them current. Each practice includes a checklist and questions to help teams gauge their work.

The practices address questions like:

• Where would a skill help address recurring problems in our work?
• What should we review before using a skill from a public repository?
• How can teams find and contribute useful skills without unnecessary approval steps?
• How do we know whether a skill reduces rework or improves code quality?

Evaluating skills also means checking the code they help produce. An agent can render different results from the same instructions, so teams still need tests and other checks to verify that generated code meets their requirements. The guidebook explains how those checks complement the guidance that teams create.

Where to start

Which practices you focus on first will depend on how your team is already using skills. Our guidebook includes a maturity self-assessment to help you evaluate your current approach and identify useful next steps. A phased plan then provides a starting point for testing these practices through a small pilot, measuring its effect on delivery, and using the findings to guide broader adoption.

We recommend using the assessment and checklists as a starting point for a conversation with your team. Identify a recurring problem, consider where clearer instructions could help, and agree on how you’ll evaluate the results. That gives you a manageable way to begin applying the guidebook to your agency’s work.

Read the AI Coding Agent Skills Guidebook.

Have questions or want help putting these practices to work in your agency? Contact our team at hello@adhoc.team.