More Prompts = Worse Code?

TL;DR

• 1

  Prompt files are technical debt too: Theo says files like agent.md, claude.md, skills, MCP configs, and dynamic system prompts all add maintenance burden, and unlike bad code, they often fail silently.

• 2

  Model-specific prompt tuning goes stale fast: A prompt setup that worked for GPT-4.1, 54, or Gemini can become actively harmful on the next model release, which is why Theo says you have to relearn how to "hold the model" every time.

• 3

  Bigger AI harnesses can make models dumber: He points to Claude Code's large prompt footprint and users who install every MCP server they can find, then wonder why half the context is gone before the agent even starts working.

• 4

  Third-party tools often beat custom setups because they retune constantly: Theo's example is Cursor, where prompt and harness work reportedly produced 10% to 30% quality gains for Opus versus other environments, plus fast fixes like a late-night Gemini 3 Pro prompt patch.

• 5

  Theo uses minimal configuration on purpose: He likes the Pie CLI because it starts under 1,000 tokens of context, compared with much heavier harnesses, and lets him see what the raw model can actually do before adding anything.

• 6

  The rule of thumb is write prompts yourself, then delete them: He warns against AI-generated markdown and boilerplate like /init files because old instructions can drag down newer models long after the team forgot they were there.