Opus 5 is the best model on the board and it talks like a consultant who bills by the syllable. Long answers, technical words, and you pay a token for every one of them. Anthropic already published the fix. It is an output style, it takes about sixty seconds, and most people have never opened the setting. Two things in the video: 1. The output style. Set once, changes how Claude talks to you forever. Mine is ELI5 plus one line of ASD-STE100 Simplified Technical English, which is the standard aircraft manufacturers use so a mechanic in any country can read a repair manual safely. 2. Skills for length. Do not hard-wire “be brief” into your style, because some days you need the long version. Put it on a switch instead. /bro rewrites the last answer in plain English. /quick 3 gives you the three things that mattered. One thing worth knowing: do not put these rules in CLAUDE.md. That file is read once at the top of a session and then slides behind you. An output style is written into the core system prompt and Claude Code keeps reminding itself about it mid-session. Same rule, two places, and only one of them still works at turn ninety. Credit where it is due: the ELI5 output style is Lydia Hallie at Anthropic. The ASD-STE100 line is Andrew Carr. The file that installs all of it is attached. Hand it to Claude Code and tell it what you want.

I made a video of a version of me that I never filmed. Commander Tom, a clone of me from the year 3192, standing in the engine room of a starship, talking. The face is mine, the voice is mine, and everything else is generated. People asked how it was made, so I recorded the whole thing and wrote the guide to go with it. Both are here. The build is three ingredients: photos of your face, a scene to put that face into, and audio. Five steps end to end. Shoot a photo set. Several angles, several expressions, and one clean forward-facing shot, which is the one that does most of the work later. Generate the scene with Nano Banana Pro inside PromptEdit. My exact scene prompt is in the guide. Swap your real face onto that scene. This is a second pass and it is the step people skip. The first render gives you a plausible person who is not quite you. My face-swap prompt is in the guide too. Make the audio. Record yourself, or clone your voice in ElevenLabs from about five minutes of recording. The audio has to exist before you animate, because the animation is driven by it. Animate with HeyGen Avatar 4, also inside PromptEdit. 720, 9:16, generate. What the run in the video actually cost, at 5 cents a credit: About 3 credits an image, so roughly 15 cents a shot 66 credits for the animation, which is $3.30 Two images plus one animation, so about $3.60 PromptEdit has no subscription, you buy credits and spend what the job costs. ElevenLabs is separate and I do not quote its pricing, so check their current rates. The guide has both prompts in copy-paste blocks, a cost table, nine things that go wrong with what to do about each, and a one-page checklist. If you build one, post it here. I want to see them.

Most people drive Claude Code entirely through prose. Eight of its built-in slash commands do the same work in a few keystrokes. /add-dir, give Claude access to a folder outside the current project /cd, move the whole session somewhere else and keep the conversation /init, write a starter CLAUDE.md you then fill with your own rules /import, bring a Codex or Gemini CLI setup across, with --dry-run to preview first /memory, open and edit what Claude remembers /ide, connect VS Code or Cursor so it can see the files you have open /keybindings, remap the keyboard shortcuts /terminal-setup, fix key combinations in terminals that need it Before you write a paragraph explaining what you want, type / and check whether the shortcut already exists. Full walkthrough in the video. Yes, that is me on a starship.

Reel Factory is a Claude Code skill that takes one raw talking-head recording and produces the finished vertical cuts you would otherwise spend an evening assembling: burned-in captions timed to the words that were actually spoken, motion-graphic cards carrying the numbers the video is arguing about, a separate platform-neutral cut that ends before any platform-specific call to action, a YouTube thumbnail, an Instagram cover frame with the hook on it, and per-platform copy that is scored against a hard threshold and rewritten for up to five rounds. Between each of those stages sits a gate that either passes or stops the run, including one that re-transcribes the rendered file to prove the neutral cut does not still say "comment TEARDOWN". The skill ends at finished files on disk. Nothing in it can post on your behalf, because nothing in it can post at all. Just hadn the link to your AI assistant, Claude, Codex, Gemini, Open Claw, Hermes Claw, and tell them to work with what they can. Claude shoudl be able to get everything going, you might have to do some work for other AIs. https://github.com/tdlyons/reel-factory

A night-shift job starts itself, does the work unsupervised, and leaves a finished draft where you will find it. Your involvement moves from the front of the job to the end of it. The part people drop when they build this is the approval step, because it feels like it defeats the purpose. It does the opposite. I am willing to let jobs run unsupervised precisely because nothing escapes without me reading it first. Remove that and you can only safely schedule trivial work, which was never the work eating your week. What is in the guide Part One: ChatGPT and Claude both run scheduled tasks now, and there is a prompt block on page four you can paste straight in. It includes the one line that makes the difference between a job that finishes and a job that sits there waiting: explicit permission to make assumptions instead of asking you a question at 4am. Part Two: The four parts every job needs, two kinds of trigger and when to use each, and four failure modes that each cost me a week. One of those is worth the download on its own: my first version decided a new article was due by checking when I last published, which meant drafts waiting for my approval never counted as done, so it wrote me a fresh one every session while three unread ones piled up. Mark when the work is finished, not when it goes live. Last page is a worksheet. Fill it in before you build and the build takes an afternoon. Try one and tell me what happened Pick the job you do at the same time every week and resent doing. Not the most important one. The one you keep putting off because it is tedious and you already know exactly how it goes. Then drop a comment with what you put on the night shift and what came back. The failures are more useful to the rest of us than the wins, so post those too.