Cole Medin uses reverse engineering and his own software-factory experiments to explain which skills still matter as agents improve (15:58).

Code generation is not the whole job

  • Delegate implementation, but stay responsible for planning, clarifying intent, and validating the result.
  • A “meat proxy” forwards a ticket to an agent and pastes its answer back without adding judgment.
  • Making colleagues review a giant “Claude said” response transfers your work rather than completing it.

Why reverse engineering is an unusually favorable task

  • Reverse Engineer Anything starts with a working application: the reference already expresses much of the target behavior.
  • New products usually begin with incomplete ideas, not a finished target that an agent can compare against.
  • Supply concrete examples where possible; otherwise build the missing reference through questions, requirements, and acceptance criteria.
  • A successful reconstruction is not proof that all engineering is solved—or that automated validation is perfectly reliable.

The 2D-to-3D example

  • Medin’s autonomous factory produced a working 2D game, but his intended visual direction was 3D.
  • Several attempts with a powerful model and elaborate workflow still missed that intent.
  • A Zelda visual reference plus repeated side-by-side comparisons brought the prototype closer to his goal.
  • He describes an early proof of concept with visible glitches, not a release-ready game.

The “twice as good” test

  • Ask: if the next model were twice as capable, would this skill matter more or less?
  • His lower-priority examples: prompt tricks, hand-built reasoning loops, and enormous generic rule files.
  • His higher-priority examples: planning, validation, references, and the tools and workflows surrounding an agent—its “harness.”
  • Turn mistakes into improvements to that workflow; do not treat stronger models as a substitute for product judgment.

“Large language models can’t read your mind.”

Context: Firecrawl sponsors the video. These are Medin’s experience-based judgments, not a controlled study or a guarantee about future employment or model reliability.