7. Reuse What Works: Skills, Agents, and System Prompts

A prompt that works once is useful. A prompt that works every time, without you having to paste it, is an asset. In OUPI, you have three ways to make a proven prompt permanent:

• Skill — a reusable instruction you save in SkillHub. When you or a colleague activates it, the instruction is applied automatically to every relevant request. No copy-pasting, no forgetting a constraint. • Agent — a more complete setup that combines a system prompt (standing instructions: tone, rules, persona, forbidden actions) with tools and data sources. Agents handle multi-step missions via Conductor. • Assistant preferences — lighter standing instructions that shape every conversation in your assistant without creating a full agent.

All three are forms of the same idea: a system prompt written once and reused many times. The difference is scope — a skill is a single instruction, an agent is a whole workflow, and preferences are personal defaults.

Not every prompt deserves to become a skill. A good candidate has three traits:

1. It solved a recurring task. You've used it (or something very close) more than twice — drafting a weekly summary, reformatting data, answering a specific type of customer question.
2. It contains the five parts that matter. Goal, context, format, constraints, and ideally an example of the expected output. A weak prompt saved as a skill just automates mediocrity.
3. It was iterated. The best skills come from conversations where you gave targeted feedback — "keep the structure, shorten section 2" — until the output was right. That final, refined instruction is what you save.

Before saving, strip anything that was specific to one run (a particular customer name, a one-time date) and keep the reusable skeleton.

When you save a skill or build an agent, you are writing a system prompt — standing instructions the model follows for the entire conversation. Write it the same way you write a great prompt:

• Start with the goal and audience: "You help HR managers draft compliant French employment letters." • Set constraints up front: "Use only the provided legal references. If a point is not covered, say so." • Show the format: paste a short example of the ideal output so the model follows it consistently. • Forbid what you don't want: hallucinated case law, informal tone, responses longer than 300 words — whatever matters.

A system prompt is not a persona game. Describing the audience and the quality standard is far more effective than writing "you are a genius lawyer."

Skills handle single instructions. But some tasks need planning, multiple steps, and tools — pulling data from a knowledge base, running calculations, then formatting a report. That is where agents and Conductor come in.

Instead of writing one enormous prompt, you define an agent with a system prompt (the standing rules) and let Conductor break the mission into steps, execute each one, use the right tools, and deliver a consolidated result. Think of the agent's system prompt as the "who you are and how you work" layer, and each step as a focused sub-prompt that inherits those rules.

Rule of thumb: if you find yourself splitting a task manually across several messages, it probably belongs in an agent.