▶ YouTube Summaries

• https://www.youtube.com/watch?v=xJaMTo2YgO8
• Original title: How I code with AI changed a lot

Theo's coding workflow has fully changed since his last video (was Cursor + plan mode + Opus). Now: GPT-5.5, the Codex harness (via T3 Code app), remote coding from a Mac Mini, lean two-sentence prompts, and single-thread-on-main instead of parallel work trees. Context built from a 5-day solo project (Lakebed, a "shitty cloud for shitty apps").

Models & harnesses

• Almost exclusively GPT-5.5 now; stopped using Claude models except quick landing pages. Generous limits on $200 plan (couldn't drop below ~6% weekly even building a full cloud).
• Defaults to the Codex harness — deliberately kept boring/minimal — used via T3 Code (an app for managing AI harnesses, not a harness itself, fully open source). Codex app is "the best bet for most people."
• Cloud Code desktop app is a third-class citizen at Anthropic; CLIs in general are worse than a good desktop app for real work.
• Warns: Anthropic now meters non-CLI subscription usage (Conductor/T3 Code/`claude -p`) at $200 cap vs $5000 via official CLI.

Remote workflow

• Wants to close the laptop and keep agents running. Codex mobile integration is decent; Codex desktop remote was broken (disappearing model picker, sticky-keys terminal lag, flaky image paste).
• T3 Code remote (Julius's work) is far better — full-speed SSH-backed terminal, image paste works, browser-hosted. Pairs well with Tailscale. One user even runs it on a Replit server from an Android tablet.
• Strongly anti-SSH-into-a-box-with-tmux; pasting images over SSH is the dealbreaker.

How he actually prompts

• Context management is the whole game. Clone reference repos and point the model at the local path instead of describing them.
• Read what the model says, not just the code/plan it outputs; steer its tone so you'll actually read it. Talk to it like convincing a colleague.
• Agents.md = a letter explaining how/why you build (no file paths, no enforced tech decisions) + a glossary of terms. Hand-written, not agent-generated. Give the model "your psychosis."
• Almost no skills installed — "just talk to the model." A "grill me" Whisper Flow binding instead of a skill.
• Prompts are usually ≤2 sentences. Don't name files — the model finds the right one better than you. Add detail only when needed; if plain language fails, fix Agents.md, don't bloat prompts.
• HTML plans (Anthropic trick) read better than markdown; once a good one exists, the model copies the format.
• Use a screenshot tool with annotation; voice-to-text (Whisper Flow) produces better/longer planning prompts.

Threads, PRs, verification

• ~100+ threads in 5 days, almost all single-prompt, run on main alone — not parallel work trees. New thread per task to avoid stale context biasing the model (re-exploring the codebase each time is fine/fast).
• Fast mode on Codex subscription (included, unlike paid Cloud Code) used liberally.
• Give the model tools to verify its own work — CLI commands, tests, or Codex computer-use (works even with the machine locked; carries from Codex app → CLI → T3 Code). ~4–5 of 50+ threads failed first try.
• PR flow: gauge complexity first; new work tree if big. Solo project needed few PRs (closed two). Reserve PRs for security/hosting-layer changes wanting extra eyes (human + Code Rabbit/Macroscope). Loop the agent against a review CLI until clean. Keep PRs from going stale — let an agent assess and close superseded branches.
• Overall theme: keep it simple; if simple fails, fix what blocks simplicity rather than adding complexity. Stop using IDE sidebars/CLIs; if you look at code more than the conversation about it, you're behind.