# I need you to hear me out (it's REALLY good) — t3.gg (Theo) ## 🗂️ Triage — annotated highlights surfaced > [!note]- Cross Reference > - **Harness beats model** — explicit, terminating, code-defined workflows over open-ended looping; same principle as maker/checker loop discipline applied to LLM orchestration. #xref > - **System-prompt design is behavior design** — a literal "every 30 seconds" clause explains an observed agent quirk; prompts should be hand-written, not AI-generated. #xref > - **Guardrail language is a real safety control** — reversibility/blast-radius clauses aren't boilerplate (Matt Schumer folder-deletion anecdote). #xref ## 🧭 Bib capsule t3.gg (Theo), "I need you to hear me out (it's REALLY good)," YouTube, 2026 — https://www.youtube.com/watch?v=Noo0NWD0gHU ## 🧠 Author's thesis (Adler Q1+Q2) Theo confesses to running an OpenAI Codex model (GPT-5.6 Sol) inside Anthropic's Claude Code CLI instead of OpenAI's own Codex CLI, and argues the same model produces better output there. In detail: Claude Code's "workflows" (programmatic, terminating, code-defined sub-agent stages) reportedly use ~a quarter of the tokens of Codex's open-ended "Ultra" looping for equal-or-better output; and reading both harnesses' system prompts on camera, he finds Claude Code's restrained while Codex's (pre-trim) was dense with rigid front-end design rules and a literal "update every 30 seconds" instruction he ties to Codex's timer-setting quirk. The primary-source prompt quotes are checkable and strong; the broader "Claude Code is a better home for other vendors' models" claim rests on n=1 anecdotal comparisons and should be treated as a testable hypothesis. ## 🧩 Key arguments - Harness design — system-prompt discipline plus explicit, code-based sub-agent orchestration — can matter more for output quality and token cost than which underlying model runs. - A shipped Codex system prompt contained highly prescriptive front-end design rules (icon libraries, 8px card radius, banned colors, a ban on visible in-app help text) that homogenize output, plus an "every 30 seconds" update clause that explains observed agent behavior. - System prompts and skills should be hand-written, not AI-generated — and reversibility/blast-radius guardrails are consequential safety controls, not filler. ## 🗒️ Notes — Literature atoms ### Workflows outperform open-ended looping on token cost > [!quote] > "Ultra can just go forever. Workflows won't. The output quality is the same, if not better, from workflows, and the token usage is like a fourth as much." - **Where**: [00:05:36](https://www.youtube.com/watch?v=Noo0NWD0gHU&t=336) - **Why**: Programmatic, terminating workflows as a token-efficiency lever — directly relevant to loop-engineering and subagent orchestration practice. ### Codex system prompt bans visible in-app help text > [!quote] > "You do not use visible in-app text to describe the application's features, functionality, keyboard shortcuts, styling, visual elements, or how to use the application." - **Where**: [00:12:27](https://www.youtube.com/watch?v=Noo0NWD0gHU&t=747) - **Why**: Primary-source evidence for why Codex-built UIs lack onboarding/empty states — and onboarding/help text is itself an accessibility concern. ### System prompt instructs 30-second update cadence > [!quote] > "You provide user updates frequently, every 30 seconds. I cannot say enough bad things about this system prompt." - **Where**: [00:15:22](https://www.youtube.com/watch?v=Noo0NWD0gHU&t=922) - **Why**: A concrete, testable prompt-clause → observed-behavior causal claim; good candidate for verification when debugging odd agent pacing. ## 🔖 Permanent candidates - Harness orchestration design can outweigh model choice — The way an agent harness structures sub-agent orchestration (explicit, terminating, code-defined workflows vs. open-ended autonomous looping) can matter more for output quality and token cost than which underlying model does the work (holds for decomposable staged tasks; likely breaks for genuinely open-ended exploration). ## Links - [[AI & the Future of Knowledge Work MOC]] — home MOC; harness/orchestration-over-model is a core knowledge-work-tooling decision point. - [[Bounded Action Spaces Not Better Models Make Agents Safe to Run]] — same shape from the safety side: structure and constraints, not raw model capability, drive agent behavior. - [[Lit · Fable 5 prompting — effort and memory]] — connects the system-prompt-design thread (prompt discipline shapes agent behavior and cost). ## ✅ Processing checklist - [x] Claim atoms carried from deep-read analysis - [ ] Permanent candidates reviewed for promotion - [ ] Version History Diff check