Most LLM-agent work focuses on tools and memory; less attention has been paid to packaging human expertise into inspectable, portable artifacts an agent can reliably invoke. COLLEAGUE.SKILL addresses this gap by automating a trace-to-skill workflow: it converts heterogeneous evidence (documents, conversation logs, decisions) into versioned skill packages with two coordinated tracks, so agents carry bounded, correctable representations of expertise rather than fragile, opaque prompts.
Key Findings
- Two-track artifact design: separates a capability track (practices, heuristics, mental models) from a bounded behavior track (communication style, interaction rules, correction history). So what: makes skills both actionable and safe to invoke, and lets teams update behavior without touching core capabilities.
- End-to-end distillation + lifecycle: automated generation, human-in-the-loop correction via natural language, versioning, rollback, and install across agent hosts. So what: reduces manual engineering effort and gives traceability and governance over agent behavior.
- Practical ecosystem: an open-source implementation with a public gallery (215 skills from 165 contributors) and a repository adoption signal (≈18.5k stars at time of writing). So what: demonstrates feasibility at community scale, not just a lab prototype.
Who it's for and tradeoffs
Great fit if you need person-grounded agents that must reflect an expert's judgments or interaction style and you have access to diverse evidence traces (logs, documents, outputs). It helps teams deploy inspectable, updatable skills across heterogeneous agent hosts and supports governance workflows.
Look elsewhere if you lack trace data, require fully formalized certifications of correctness, or need continuous real-time learning from production signals—COLLEAGUE.SKILL depends on the quality and representativeness of input traces and focuses on packaging and governance rather than online continual learning. Also consider privacy and IP constraints: distilling personal traces into distributable packages requires careful handling of sensitive content.
Where it fits
This paper sits at the intersection of LLM agent design, persona/memory research, and AI tool governance: it reframes personalization as a packaging problem (portable, versioned skills) and offers a practical workflow for teams that want auditable, correctable person-grounded capabilities rather than ad-hoc prompt hacks.