The paper presents GEIS (Generation-Evaluation-Improvement loop of agent Skills), a framework for Wikipedia-style long-form article generation that replaces fixed multi-agent pipelines with named, declarative skills. Implemented in Tasi Harness, the system composes specialized modules for writing, evidence collection, and evaluation to enable inspectable and iterative improvement.
- GEIS implements a core writing skill following a Request, Plan, Draft, Audit, Refine, and Deliver sequence.
- A pairwise evaluation skill generates structured quality reports using PDF-aware assessment.
- An improvement skill maps recurrent findings into permanent patches for the writing skill.
- In a 20-topic experiment, GEIS improved over the Tasi Harness default writer by 8.0 points on a 100-point PDF quality rubric.
- The patched writing skill raised the average score from 82.90 to 86.95, with 17 out of 20 topics showing improvement.
The authors argue that reframing long-form generation as an evaluation-guided loop allows for modular and inspectable workflows compared to entangled prompt-based systems.