Louis-François Bouchard defines common weaknesses in generated writing, including generic language, repetition, unsupported claims and a mismatch with the authors tone. He describes a benchmarkA benchmark is a defined test or collection of tasks used to compare a system's performance under specified conditions. built around his own script-writing style.
The benchmark uses multiple drafts for each task and several model judgesEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria. to score tone, writing quality, instruction following and research accuracy. Bouchard discusses judge preferences, human pairwise comparisons and the costs of changing prompts or scoring rules.
Bouchard recommends detailed examplesContext engineering designs how relevant instructions, evidence, state and tool information are assembled for a model's task., limited review passes and a final human reviewHuman judgment in AI is the accountable interpretation and decision-making people contribute when setting goals, evaluating evidence, and managing consequences.. He cautions that scores based on his voice do not establish a universally best writing model.
Watch on YouTube
