Everything AI makes sounds the same. That's the slop problem, and it starts with what you feed it, not with the model.
In this episode I set up two brand new accounts, Claude and ChatGPT, from scratch. Same three layers in both: Account, Project, and Context. Then I gave both the same real RFP for a nonprofit website rebuild and asked for a proposal. They landed five hundred dollars apart on a hundred thousand dollar estimate.
Who you are and how you work, in every chat. Paste into Claude's personal preferences or ChatGPT's Custom Instructions.
Download for WordMarkdown versionOne area of your work and what it needs. Overrides your account level. Paste into a Claude or ChatGPT Project.
Download for WordMarkdown versionContext Rot: How Increasing Input Tokens Impacts LLM Performance. Hong, Troynikov, Huber. Chroma Research, 2025. Eighteen models tested; every one performed worse as input grew.
https://www.trychroma.com/research/context-rot
NoLiMa: Long-Context Evaluation Beyond Literal Matching. Modarressi et al. LMU Munich and Adobe Research. ICML 2025. Eleven of thirteen models fell below half their baseline at 32,000 tokens when the question didn't share wording with the answer.
https://arxiv.org/abs/2502.05167
The RFP: Hunger Task Force website redesign, posted publicly and closed in April 2026. Drafted here as a working exercise.
https://www.hungertaskforce.org/2026/03/18/website-redesign-rfp/