▶ YouTube Summaries

• https://www.youtube.com/watch?v=WnBx1Vi7M6w
• Original title: Your prompts are tech debt.

Theo reacts to Sean Goedecke's argument that prompts — AGENTS.md, CLAUDE.md, skills, MCP configs, system prompts — are a worse form of technical debt than code. Code debt is loud (you feel it when editing); prompt debt decays silently. A prompt carefully tuned for one model can go stale or actively harmful on the next model upgrade, and you may never notice because model behaviour is hard to measure. His takeaway: write prompts yourself, keep them minimal and factual, delete aggressively, and lean on third-party harness teams who re-tune per model release.

All code is debt — and so are prompts

Every line of code adds maintenance burden; sensible engineers write as little as possible (Theo: proud of deleting more than he added at Twitch). Neither humans nor AI can hold a million-line codebase in context — both rely on grep-style search, not full-context loading. He calls out Cursor's early bet that ever-bigger context windows would let AI load the whole codebase; Claude Code's "just grep for it" approach won instead.

Prompts as a new, silent debt class

Modern projects accumulate prompt files, skills, dynamically built system prompts (T3 Chat tweaks its prompt per model/tool, "ungemini"-ing Gemini). Tiny prompt tweaks unlock big performance shifts — the Cursor harness reportedly gives Opus a 10–30% quality bump over Claude Code, mostly via system prompt and per-model A/B testing. System prompts aren't secrets worth guarding, but the engineering in them is real.

Why heavy customization backfires

Prompt tuning is model-specific; labs re-tune tool descriptions and prompts for every release (codex with 55 mis-behaved until descriptions were fixed). A prompt crafted in January can be harmful by February — silently. Avoid blindly installing MCP servers and skills: they bloat context (Claude Code system prompt ~65K tokens loaded, ~12K minimal; the Pi CLI starts under 1K). Models will use tools just because they're present.

Practical advice

Pick a third-party-maintained tool (Claude Code, Codex, Cursor, T3 Code), keep it near-stock, disable MCPs/skills by default, and piggyback on the vendors' per-model tuning. Avoid behaviour-steering fluff ("think step by step", "I'll tip you $200"); keep AGENTS.md to concrete project facts. Don't let AI generate your prompt files (the claude `/init` output is slop). Theo's own global AGENTS.md is three lines and he's considering deleting it. Action item: audit your markdown/prompt files — the T3 code AGENTS.md was untouched for two months across a fully rewritten codebase and still contained outdated, possibly harmful framing.