Jumping in here to add my support for Skills becoming a first-class feature in Copilot Studio.
For certain use cases, I need to give an agent clear, structured answer guidance, such as a template for how to fill something out using content it already has access to. The problem is that embedding full templates directly in the system prompt quickly exhausts the 8,000 character instruction limit, and with two or three distinct use cases requiring detailed guidance, that limit becomes a hard blocker.
I've found two workarounds that function reasonably well in practice:
1. Store a skill.md file in OneDrive and use the Read OneDrive File tool, referencing it in the instructions so the agent reads the file before proceeding.
2. Decompose the agent into sub-agents with dedicated, focused instruction sets (assuming 8k chars is sufficient per sub-agent).
Both approaches work, but they're exactly that: workarounds. The experience is fragile and harder to maintain at scale.
A more robust native solution would look something like this:
a) Skill management at the agent level, modelled after how knowledge sources already work today. Skills should be defined separately from the system prompt and invoked conditionally based on context, not loaded wholesale into every turn.
b) An organisation-wide skill marketplace where teams can publish and centrally maintain reusable skill files. The obvious example: a marketing team hosting brand voice and content guidelines as a managed skill, available to any agent across the organisation that generates customer-facing or internal content. The same model applies to legal disclaimer templates, support response frameworks, HR policy guidance, and so on.
This kind of architecture would make Skills genuinely composable and governable at enterprise scale, rather than a workaround you bolt on per agent. Would love to see it on the roadmap.
