GEIS: A Generation–Evaluation–Improvement Loop of Agent Skills for Long-Form Article Generation

Long-form generation tasks—such as writing Wikipedia-style articles—are challenging because they require managing long contexts, complex instructions, and lengthy outputs. Existing multi-agent systems often rely on rigid, entangled prompts that are difficult to debug or refine. The authors introduce GEIS (Generation–Evaluation–Improvement), a framework that treats writing, retrieval, diagramming, and evaluation as independent, modular skills.

GEIS centers on an article-writer skill that follows a strict six-stage process: Request, Plan, Draft, Audit, Refine, and Deliver. By separating these capabilities, the system allows for explicit quality gates (Audit) and enables the article-writer-improving skill to map recurrent evaluation feedback into permanent, actionable patches for the writing process itself.

The authors evaluated GEIS across 20 Wikipedia Featured Article topics, comparing it against a default harness writer and the STORM framework. GEIS demonstrated superior performance, scoring 8.0 points higher than the default writer on a 100-point PDF quality rubric. When compared to STORM, GEIS showed improvements in structural and content quality.

In a dedicated improvement experiment, the authors used the system's feedback loop to generate permanent patches for the writing skill. This iterative process raised the average quality score from 82.90 to 86.95, with 17 out of 20 topics showing measurable improvement, primarily driven by gains in content quality.

This research shifts the paradigm of long-form generation from "black-box" prompt engineering toward an inspectable, engineering-oriented lifecycle. By framing writing as a series of auditable skills rather than a monolithic task, developers can systematically diagnose failures and evolve agent capabilities over time. This approach makes complex, multi-step generation tasks more reliable and easier to maintain in professional or technical documentation workflows.