Versioning the behaviours and workflows I use to engineer with AI.

Source on GitHub License: Apache-2.0

A collection of specialised AI agents

Why I built it

An AI coding tool can generate code. The harder and more interesting problem is making the collaboration consistent: when should the agent ask questions before acting? How should it diagnose a failure? Which perspective should challenge an architecture, and what should the result look like?

Those decisions are engineering knowledge. I built Agent Skills to keep that knowledge as ordinary, inspectable files rather than losing it in one-off chat sessions or hiding it inside a large framework. The collection lives in Git, so the instructions can be reviewed, refined, shared across projects, and rolled back when an experiment makes them worse.

Skills and agents solve different problems

The repository contains two related kinds of reusable behaviour:

  • Skills describe focused workflows that an agent loads when a task matches. The current collection includes idea-interview, for discovering and sharpening a point of view, and blog-writer, for turning the resulting material into a standalone technical article.
  • Agents establish a specialist role, its boundaries, and its expected output. They include an architectural critic, product and technical product owners, a reliability engineer, a troubleshooter, a podcast host, and a precision reasoning persona.

The distinction matters. A role answers who should think about this? A skill answers what repeatable method should they follow? Keeping the two small and composable makes them useful beyond the project where an idea first emerged.

One workflow used to develop this blog is deliberately simple:

rough idea
  → idea-interview
  → sharpened thesis, examples, and objections
  → blog-writer
  → Jekyll article

The value is not the diagram. It is that each transition has explicit expectations, so the human remains involved in the decisions while the agent handles more of the mechanical work.

AI-enabled engineering is more than prompting

I treat agent behaviour as part of the engineering environment. Useful instructions need clear triggers, narrow responsibilities, and boundaries that stop a specialist from silently becoming a general-purpose decision maker. They also need maintenance: prompts drift, assumptions leak out of their original projects, and a workflow that sounds good on paper may be awkward in real use.

Agent Skills is where I work on those problems in public. It reflects the same principles I apply to software: make responsibilities explicit, prefer small composable pieces, keep the source readable, and use version control to preserve the reasoning behind change.

Plain files, linked into local tools

The implementation stays intentionally close to the metal:

.
├── agents/     # specialist personas
├── skills/     # reusable workflows, one SKILL.md per skill
└── scripts/    # local linking and discovery helpers

The linker creates symlinks into supported runtime directories that already exist, including ~/.claude/skills, ~/.claude/agents, ~/.agents/skills, and ~/.config/opencode/agents. Because the links point back to the repository, an edit is immediately available to the local tools without copying prompt files between them.

To try the collection:

git clone https://github.com/abbyssoul/agent-skills.git
cd agent-skills
./scripts/link-skills.sh
./scripts/list-skills.sh

The repository also includes ./scripts/unlink-skills.sh to remove symlinks that point back to the checkout.