Orbit Alchemist

Inspiration

AI agents are powerful, but they usually enter a repository with no project memory. They do not know the testing style, CI rules, architecture, ownership, review expectations, or hidden team conventions. Humans learn this context over time. AI agents need a way to extract it quickly and preserve it for future work.

Orbit Alchemist was inspired by that cold-start problem: how can GitLab Orbit’s structured repository context become reusable expertise for future AI agents?

What it does

Orbit Alchemist is a GitLab Duo agent that turns repository context into reusable AI artifacts.

It uses GitLab Orbit and repository files to analyze a project, then generates:

• AGENTS.md: a compact operating guide for future AI agents
• SKILL.md playbooks: task-specific instructions based on real repository evidence

For example, instead of telling an agent to “follow best practices,” Orbit Alchemist can generate a skill like “Add a New glab CLI Command” with exact package layout, test patterns, helper functions, docs generation steps, and common review mistakes.

How I built it

I built Orbit Alchemist as a prompt-driven GitLab Duo Agent Platform project.

The project includes:

• a custom agent YAML config
• a large MASTERPROMPT.md
• an AGENTS.md operating guide
• an orbit-alchemist-analyze skill
• a skill-creator skill

The agent supports two usage modes:

• visible mode, where an orbit-alchemist/ folder is added to the target repo
• direct catalog mode, where the embedded master prompt is used as a fallback

The workflow starts with Orbit graph context, then scans repository files, distills conventions, and generates evidence-cited artifacts.

Challenges I ran into

The biggest challenge was making the agent reliable instead of generic. Early versions could answer from repository files but did not always prove that they had loaded Orbit Alchemist’s own instructions first.

I solved that with a bootstrap flow: the agent must state whether it loaded the visible orbit-alchemist/ files or used the embedded fallback workflow.

Another challenge was GitLab prompt validation. The system prompt had to avoid unsafe patterns like HTML comments, angle-bracket placeholders, and hidden Unicode characters.

Accomplishments that I'm proud of

I’m proud that Orbit Alchemist produces repository-specific knowledge instead of generic advice.

The generated skills cite real files, real review rules, and real project patterns. A good generated skill would not work copy-pasted into another repository, and that specificity is the point.

I’m also proud of the two-mode setup: teams can either keep the agent instructions visible in the repo or use the catalog agent directly with the embedded fallback prompt.

What I learned

I learned that repository intelligence is not just documentation. It is operational memory.

For AI agents, the most useful context is not a long wiki page. It is compressed, evidence-backed guidance that tells the agent exactly how to work in this specific repo.

I also learned that skills become much more useful when they can include references, examples, scripts, and assets when needed, instead of only a single Markdown file.

What's next for Orbit Alchemist

Next, I want to make Orbit Alchemist generate merge requests with the produced AGENTS.md and skills, not only comments.

I also want to add stronger validation for generated skills, richer Orbit queries, better detection of repo-specific workflows, and automatic refresh when major repository changes happen.

Long term, Orbit Alchemist could become a reusable onboarding layer for AI agents across GitLab projects.

Built With

• agents.md
• ci/cd
• gitlabduoagentplatform
• gitlaborbit
• markdown
• skill.md
• yaml