Video summary

More Prompts = Worse Code?

Main summary

Key takeaways

Technology

Tech Debt vs. “Prompt Technical Debt” (central idea)

  • Traditional technical debt is already common and dangerous in software:
    • complexity grows,
    • systems become harder to understand,
    • future changes get riskier.
  • With AI coding agents, a new form of technical debt can emerge: prompt technical debt.
    • This refers to system prompts / agent.md / tool descriptions / skills / MCP setups stored in repos or inside an agent harness.
    • These assets can become outdated or actively harmful as models and tooling evolve.
  • Unlike normal code rot (which often fails visibly), prompt decay can be silent:
    • A prompt that “used to work” may later degrade performance or cause incorrect tool behavior after a model update—without obvious errors.

Why prompts become debt

  • Prompts strongly affect behavior and quality.
    • Even small changes to an LLM prompt can unlock large performance improvements.
  • System prompts are not portable across model versions.
    • A prompt tuned for one model release may underperform or cause regressions on the next.
  • Tool/workflow configuration is also prompting.
    • Adding route loops, skills, or MCP servers changes what the model sees.
    • Often this flows through the system prompt, tool registry, or similar mechanisms—effectively altering behavior.
  • Over-installing tools/plugins wastes context and biases tool use.
    • MCP servers can consume large portions of context.
    • Models may use available tools even when you don’t want them to.

Real-world examples / claims mentioned

  • The speaker flags agent.md in the T3 code repo as outdated:
    • e.g., “early whip” / “sweeping changes encouraged” language,
    • with concern it may lead models to push unnecessary changes.
  • Prompt tuning examples:
    • Switching between coding environments (Cursor vs. Claude Code vs. “official” integrations) can yield measurable quality differences, largely attributed to system prompting.
    • The Cursor team reportedly does AB testing and prompt benchmarking.
    • One anecdote mentions Gemini 3 Pro requiring an “un-Gemini-ify” style blurb to prevent undesirable tool usage.
  • Failure mode explanation:
    • Even if a harness works today, every model upgrade can turn a functional prompt into non-functional behavior.
    • The author questions leaving outdated “behavior steering” language in agent prompts.

Guidance / best practices advocated

  • Prefer third-party maintained tools (e.g., Cursor, Claude Code, Codex, Copilot, T3 code) and keep them as unconfigured as possible.
    • This helps you benefit from vendors’ ongoing prompt/tool maintenance.
  • If you do maintain prompts:
    • Audit markdown/system prompt files regularly, especially those untouched for months.
    • Avoid bloated agent.md files:
      • don’t let models write huge pages of barely reviewed text into prompts,
      • same philosophy as “don’t write lots of low-quality code.”
    • Avoid overly directive / behavior-steering lines that are generic or potentially misleading:
      • step-by-step instructions,
      • tipping threats/rewards,
      • “don’t make mistakes.”
    • Keep prompts limited to concrete project facts and clear intent.
    • Delete prompts or prompt-generators when you can, especially if you used AI to author/generated prompt markdown that is now stale.

Mentioned tooling: “pie” (minimal context as a mitigation strategy)

  • The video highlights the pie coding agent/CLI:
    • It starts very minimal (speaker claims < 1,000 tokens of context initially), compared with larger harnesses.
    • Philosophy: begin with the smallest possible setup, then add tools only when needed.
  • This “Unix philosophy / minimalism” approach is presented as a way to reduce prompt/context blow and regressions.

Takeaway recommendation (what to do next)

  • Treat prompt files like production code: prompts can rot, and model upgrades accelerate that risk.
  • Do a prompt-system audit:
    • system prompts,
    • agent.md,
    • tool lists,
    • MCP/skills,
    • remove stale/outdated steering instructions.
  • Minimize custom harness complexity; let third-party teams do the prompt maintenance.

Main speakers / sources (as referenced)

  • Shawn Godexx (main referenced author)
    • speaker endorses his articles and argues for “prompt technical debt” framing
  • Michael Truel (CEO of Cursor)
    • referenced for context about AI/coding approaches
  • Joel
    • referenced in agreement: “agent.md is the best thing that came out of open claw”
  • Bad Logic / Mario
    • author of the pie project

Original video