Contradictions

Where speakers genuinely disagree — claim pairs surfaced from the knowledge graph, then verified by a judge panel before they appear here. Each side links to the moment it was said.

Context minimalism — giving the model only minimal system prompt and minimal tools — is the right approach with modern models, replacing context engineering and prompt engineering.

“with the models of today, you don't do any of this. You give it the minimal possible system prompt, the minimal possible tools, and then you let the model figure it out”

— Boris Cherny, Reflecting on a year of Claude Code @ 26:56

The context provided to Claude in a routine is the ceiling of how successful Claude will be.

“whatever context Claude has, that's uh the ceiling of how successful Claude will be”

— Maya, Build a proactive agent workflow with Claude Code @ 19:10

why in tensionBoth address how much context/information to give the model, but Claim A advocates minimalism while Claim B implies that richer, more carefully engineered context raises the ceiling of performance, pulling practitioners in opposite directions on how much to invest in context construction.

Skills should be discovered through iterative exercises and muscles rather than designed upfront, because the SOP emerges from doing the work.

“I don't build skills, I discover them, right? I I don't think of it as like I already know the way that you're going to do this. Like I don't know the SOP. like figure out the SOP, standard operating procedure, like figure out how to do the thing”

— Alex Damis, To Give an Agent a Workspace: Laying a Foundation for Messy Forever Work @ 49:20

Skills should be kept to a small focused set (5–10) and made progressively better rather than expanded broadly, because tooling changes too fast to maintain a wide skill surface.

“we centralize our leverage around five to 10 skills. we don't go super wide on skills preferring to make the existing skills better”

— Ryan Lopopolo, Harness Engineering: How to Build Software When Humans Steer, Agents Execute — Ryan Lopopolo, OpenAI @ 22:26

why in tensionBoth address how to approach building/managing agent skills, but A recommends emergent discovery through iteration rather than upfront design, while B recommends constraining to a small deliberate set—a practitioner must choose between letting skills proliferate organically versus strictly limiting scope, pulling in opposite directions on skill portfolio strategy.