AI Agent Studio: reusable skills for multi-agent systems
I built plug-in skills (bias handling, voice and tone, output formatting) and a visual/UI evaluator, so multi-agent systems inherit Oracle’s behavioral rules and generated pages can be scored against the spec.
Roles & responsibilities
Prompt engineering: bias, voice and tone, and formatting skills
Systems design: layering into core, on-demand and code
Tooling: the visual/UI evaluator
Collaboration
Platform engineers
Information development
Tools
Figma
Codex
Claude
Internal environment
Impact
Published internally as the skill set for teams on AI Agent Studio
The decision
Keep always-on skills thin, load detail on demand, and move checkable rules into code.
Why. Every always-on line costs tokens on every request. Layering keeps the cost small and gives each rule something that can check it.
The problem
Teams building multi-agent systems on AI Agent Studio each re-derived Oracle’s voice, formatted output their own way, and had no standard answer for biased requests. The guidelines lived in documents, and documents do not run inside an agent. Once pages were generated, there was also no consistent way to check them against the spec.
The skills
I turned the internal guidelines into plug-in skills any team can add to an agent. My focus was standardization: the same voice, formatting and bias handling in every agent.
Bias topic
Keeps bias out of agent output even when the user’s own framing is biased, applied consistently by orchestrators and specialists.
Voice and tone skill
Oracle’s voice as an executable skill, so any agent that loads it sounds like the same product.
Output formatting skill
When to use bold, lists, tables and headings, written as decision rules. No more three agents, three styles.
Visual/UI evaluator
A skill and Chrome extension that scores a generated page against the benchmark and says how to improve it. See below.
How they layer
Always-on skills sit in every request, so their size is a real cost. I designed the skills in layers: a thin core that is always on, fuller guidance loaded only when triggered, and code passes for rules that need no model at all.
Visual/UI evaluator
A skill, also built as a Chrome extension, that checks an internally generated page against the benchmark. It ingests the internal specs, evaluates the page, then returns a score and recommendations on how to improve it.
- Reads the specs. Ingests the internal specs it grades against.
- Looks at the real page. Runs on a generated page in the browser.
- Gives a score. One number against the benchmark.
- Says what to fix. Recommendations, tied to where the issue is on the page.
What I learned
A skill has a budget. Every line costs tokens on every request, so write only what changes model behavior and push everything deterministic into code. And a standard is only real once something can check it.