Your prompts are tech debt.

The presenter argues that the prompt files developers now litter through repositories—AGENTS.md, CLAUDE.md, bespoke system instructions—constitute a new, insidious form of technical debt. Unlike traditional code, which predictably breaks with stack traces or compile errors, prompt debt decays silently. A prompt tuned perfectly for Claude 3.5 Sonnet in January might actively sabotage Claude 4's reasoning in February because model behavior shifts under the same instructions. The implication is not theoretical; it reproduces regularly in tools like Claude Code, where Anthropic’s own system prompt tweaks occasionally trigger regressions that users only notice through degraded output quality.

The practical recommendation is deliberately minimalist: rely on third-party coding harnesses (Cursor, Claude Code, Codex) kept as close to stock configuration as possible. These vendors employ full-time engineers who A/B test system prompts against each model release—work individual teams cannot replicate at scale. Customization should be additive only when a concrete, measurable problem exists, and removed the moment it is no longer demonstrably necessary. Every installed MCP server, skill file, or plugin pollutes context and increases the surface area for silent failure.

When writing project-level prompt files, the advice is to limit them to specific, factual information about the codebase—file locations, naming conventions, architectural constraints—rather than behavioral steering (“you are a senior engineer,” “think step by step”). The former ages gracefully; the latter becomes actively harmful. An audit of existing prompt files is recommended immediately, precisely because the decay is invisible until it manifests as subtly worse code generation that engineers may misattribute to model quality.