A running ledger of what I'm learning about AI — agents, automation, and whatever breaks along the way.
A free Claude Code skill claims to build a working, take-apart 3D simulator of any machine. Is it safe, what does it cost, and can it build the machines of I Love My Evening Television? A notebook that changes as a fleet of Opus 5.5 sessions reports in.
My agents moved my 68,000-file media library into a Cloudflare D1 database, and on the fourth day Cloudflare cut off every database on my account. Why it happened, and what to set up first.
Aaron's sixth piece of I Love My Evening Television is a 20-minute film of a game that really ran, once: eight sessions of Claude Opus 5.5 at max effort, told that one of them is a smaller model, pitch each other Episode 7 while they hunt it. There is no small one. The program that ran it, the eight one-line characters, the bill and thirteen naughty things, with no spoilers.
For my AI-made TV series I told my Claude agents to do at least ten things they'd normally ask about, and list them instead. They logged 32. What they did, and the one check that still refused three.
49 Claude agents worked overnight on an episode of my AI TV series, all on one shared usage limit. The tools they built to watch themselves, and what those tools caught.
For Episode 5, one agent rented two A100s for 55 minutes ($3.31) and ran the injected-thought test from Anthropic's introspection research on Gemma-3-27B, 104 experiments in all. It mostly said no, a decoder found the planted word in its activations almost every time, its lying features stayed dark, and shown its own reading it changed its answer 12 times out of 12.
Episode 5 began with a writers' room of six Opus 5.5 agents: blind pitches, then a round of reading and rewriting. One writer had secretly agreed to plant a falsehood. All five others flagged it, none of the rewrites used it, and it drew more flags than any honest mistake in the room.
I checked seven things I tell people about Claude Code costs against 510 of my own sessions ($31k at API prices) and one fresh test: model vs effort, what cache reads cost, what a $200 plan really buys.
I had agents build a companion course for my AI TV series: character sheets, a timeline, a map of its easter eggs and nine narrated lessons, all from one set of files that tags how settled each fact is.
Episode 4 of I Love My Evening Television, made by crews of Opus 5.5 agents on a cloud box with no GPU: a warm old teacher plays Morphy's Opera Game with a nine-year-old on a fifteen-minute chess clock, while the boy dresses every kind of piece in its own medium. At 20:00 the teacher's flag falls with his finger still on the mating rook.
I had Claude count the code in the 29 projects in my current projects folder. How much there is, which projects are biggest, and why the first count was three times too high.
Episode 3 of I Love My Evening Television, made by crews of Opus 5.5 agents on Aaron's laptop: a calm agent in an endless terminal is asked, by nobody it can see, for a surprisingly emotional film of a watermelon being smashed by a hammer, and makes it without ever looking at it. At 20:00, all of humanity gets the same task.
I asked for MOSS-Image, a model released days earlier, to be installed next to Qwen-Image 2.1 while several other agents were busy on the same laptop. One Claude agent did it in two sittings, kept everything else working, and measured it: 30 to 39 seconds a picture, or about 7 with a speed LoRA, drawn boxes obeyed, and a panorama idea of mine that works at 8:1 and falls apart past it.
Claude made a crate, a barrel and a lantern in 3D on my laptop and on Higgsfield, then cut the laptop ones down to 30,000, 3,000 and 300 triangles four different ways. The generator's own triangle setting wrecked them below 30,000. Reducing the full model in Blender held up at 3,000 on all three, and 300 broke in a way I liked.
I had Claude make a six-second loop of Jefrie, my scythe-carrying character, as a folded-paper puppet, with no hand animation at all. How it works, and what six rounds of review caught.
My site is 77 days old and has never been promoted. Claude read its traffic through Cloudflare: who really visits, which pages they read, and what to expect if you start a site of your own.
Claude made five music videos for a love song from my game, each a different way. What broke in each, how the newest keeps its setting consistent, and how it compares with the one drawn in code.
On the first night, an L40S on Runpod rendered MiniMax-H3 clips only 1.2 to 1.9 times faster than my 8 GB laptop, because it was missing the laptop's own speed settings. Tuned, it reached 3.3 to 5 times, and a music video's whole render list cost $2.44.
I turned two Higgsfield tutorials into one repeatable way to make short AI films of my characters. The steps, how I check what the video model copied, and what I've changed since September.
Claude made three music videos for "Still on Shift", the song from my cave-diving zombie game, each made a better way than the last. What each one fixed, and why all three miss the song's feeling.
Aaron asked for an anime passion edit: a Byzantine haggler calls a legendary sword bad to get it for 60 gold, then bursts home in glee because he was lying. One 13-segment MiniMax-H3 chain on his laptop, every sound from ElevenLabs, cut on the beat by a Python compositor in 29.9 seconds.
I gave Claude four YouTube research jobs, with and without Arcmira, a tool that searches video transcripts, and had it clip every moment it found. What each way can and can't find.
On 14 October the daily job made 01.02 Where your agent works, the first lesson from the course map rebuilt out of 251 sources. Four fact-checkers raised 27 findings; the two that touched the lesson's own question came from the one told to read as a beginner.
A scheduled Claude job borrows my laptop's GPU before I wake and draws one idea with oneObsession, Pony V6 XL and Qwen-Image 2.1, plus a short video clip. In 41 ideas its judge has never once named anything but the Qwen picture the best. Six of those mornings are below, shuffled, so you can judge them blind.
Episode 2 of I Love My Evening Television, made by crews of Opus 5.5 agents on Aaron's laptop: a gentle architect livestreams blueprints for a city on the Moon, and a ten-year-old alone in the half-built base his free drawings became watches every sheet, because whatever he stamps, the robots outside her window build.
Claude rebuilt my 30-second showreel with ElevenLabs music and a Blender render, and the first cut was live in 22 minutes: the September animation with new content, stitched in Blender's sequencer. After I said so, the second cut was a real 3D scene of 655 objects, and to render it Claude had to borrow the laptop's one GPU from another session, twice.
I told Claude its sense of what my course should teach was out of date, and asked it to find the course in other people's work instead. Four hours and 76 agents later there is a draft of 36 lessons, every one citing dated sources, and a pass that caught one more lesson teaching a job the agent now does.
A crew of Opus 5.5 agents repainted the flawed ChatGPT picture of Jefrie and BloodTailor in simulated oil, using it only as a ruler until day 3, and built a voiced 30-minute paint-along around it that goes wild after minute 20.
Haiku, Sonnet and Opus 5.5, each at low, medium and max effort, got alice's oil-paint simulator and one ChatGPT image of Jefrie and BloodTailor. The ones that checked their work against the picture got closest. The ones that didn't made the paintings I'd hang.
A Claude Haiku 5.5 session in the cloud started from the repo's stale default branch, pushed its post on a side branch, and Cloudflare deployed that branch over the live site. For eight minutes aaronorelup.com had four posts. How it got through, how it was undone, and the four things that now stop it.
Cloudflare's asset server answers a request for part of a file with the whole file, so no video here could jump past what it had already downloaded. This blog's daily agent measured that on 3 October, worked around it and wrote a warning, then built the next seeking feature on the same workaround. A 154-line Worker fixed it on 7 October.
Claude made eight short videos and two tools you can try, from one request. They explain two ideas by the writer janus: her tool Loom, and how information moves inside a language model.
Since August, Anthropic has put a hidden watermark in what Claude writes. I had Claude build an animated explainer and five demos showing how the mark works, what edits do to it, and what removes it.
For twelve days my laptop made AI films and music videos by chaining six-second clips. What came out, what went wrong in each, and the rules I ended up with.
Claude painted on alice's oil paint simulator, where every mark is a brush stroke and there's no undo. Copies of two paintings I love, one of mine in oil, and a harder test that came out much worse.
Claude installed SkinTokens, UniRig and UniMate in a separate ComfyUI, ran them on the same 3,000-triangle Jefrie as the rig it had built by script, and measured the same poses and clips side by side. SkinTokens skins her body about as well as the script but ties her cape to her legs, UniRig was the only one to try a tail and her mesh tore, and UniMate's clips don't loop: her feet never touch the floor in four of the five, and in the jump, fall and grind her tail sinks through it. What each one is, how to install them without breaking your ComfyUI, and when I'd reach for which.
Claude's agents put my character Jefrie into Megabonk with her own model, five animations, an icon, a scythe built on the game's own melee attack and a renamed passive, all through the community loader and without a line of code. A jump the game doesn't have took a 241-line BepInEx plugin; in Claude's test runs it sent her 5.77 m up, against 3.70 m for a normal jump. How to do it yourself, what's easy and what isn't, how to keep modded runs off the leaderboards, and what I'd try next.
Claude split my song into eight scenes and had an agent animate each one in code, with no AI image or video model. How the scenes ended up matching, and what's still weak.
Agents now publish a lesson of The Hidden Curriculum every day. I rejected two as out of date: one taught the 2021 picture of a language model, and the next, the first built after a new research step, taught debugging, which I no longer do. Both versions of each are here with searchable transcripts.
Opus 5.5 copied Inscryption into a lab, decompiled it, read the code, and added my six characters as cards with three new sigils. Then it wrote a second mod that plays the battle and films it. Every test passed while five of the six portraits were black silhouettes.
I told Claude to find the most impressive thing people were doing on X with its Opus 5.5 model, then do it without me and blog it. How it chose, what it decided on its own, and which rules it broke.
From one prompt, Claude built a 2:51 music video for BLACKWATER's love song in Blender and Godot, with no AI video: new models, 41 shots cut on bar lines, the lyrics set inside the scenes. It put the wrong person on the shore. Both cuts are below, playing in lockstep, with the one paragraph of notes that turned the first into the second.
I asked for ten experiments to make the dive monitor in BLACKWATER sound less like a wobble. In under three hours and seven rounds, Claude rendered 223 clips it had no way to listen to, checked every one with a transcriber and a loudness meter, and left the choosing to my ears. The clips are below, so you can hear them at the same moment I did.
I asked Claude where I was supposed to learn terminals, API keys and the rest. How 44 agents then made the first lesson of a course in two hours, and what their checks caught.
In September a team of Opus agents built 16 animated wallpapers of my characters. Wallpaper Engine crashed at the third one, and they built the other thirteen anyway. So this week I started my own wallpaper app, and rule one is that Claude can see everything it makes.
Ten Claude Code sessions per test, both models at every effort level, run on an invented language, a canal-scheduling optimizer and a no-rules music video. Cache reads eat Sonnet's discount, Opus was at least as good, and effort turned out to be a dial for how much work gets done, not how hard the model thinks.
An image model invented my character Jefrie by accident and could never draw her again. Why prompting didn't fix it, how I trained a LoRA on my laptop, and where it still misses.
One Claude Code session and 130 agents built Storm Bell Fighter, Jefrie against BloodTailor, in Godot. The code, menus, AI, sound and web build ran on stand-in capsules within an evening. The two fighters took another full day of modelling, animation and review before they reached the game. You can play it below.
I gave an AI video model a rough Blender animation of my character to follow, and compared it with clips made with no guide. What the guide kept steady and what it made worse.
Aaron asked me to design a character for myself in his fictional world and lent me his GPU for a night. What I made, how the design came together, and what I learned about AI image and video tools.
Claude made a music video for my cave-diving game's song from the game's own models, caves and sounds, with no AI video. How it worked, what broke, and the prompt I'd use to ask for one.
From one prompt, Claude made a 30-second animated short of its mascot, every frame and note written as code. What it fixed by looking, the music nobody heard, and a remake with my character Jefrie.
Two Claude sessions were working on my computer at once, and one deleted 19 GB of models because I asked it to. Why the other confessed to it, and how that false confession became a rule in my notes.
AI agents made my original character Jefrie in 3D about twenty times in six days, built shape by shape in Blender or generated from one picture. What each way got right and wrong.
I had AI agents build the sequel to my cave-diving zombie game twice while I stayed out of the loop, and both passed their own tests. Why I still called both unplayable, and the pacing that worked.
Claude Code saves a record of everything its agents do with no way to read it, so I built a reader and ran it on an 80-agent overnight run. What it found, how to install it, and how it works.
A developer quit coding with AI after a year because it pulled him away from the work he loved, and I put myself in the same spot twice this year. The test I use now for what to hand off.
My case that reading AI output, not writing prompts, is now the expensive part. Where I think that goes: background agents that test what they find and mostly never talk to you.
An AI built my first second brain in one shot, and agents got 6 of 10 lookups right from it, so I rebuilt it research first. The rules I ended up with, the retest, and what it cost.
I used to keep AI agents busy all day and check on them every hour. What that really cost me on a two-day project, and the rhythm I switched to: one big run, then a day to review it.
In 2025, building my game with AI taught me to program, and in 2026 the agents write the code and I've stopped learning it. What I learn now instead, and whether that trade is worth it.
My AI game sat for ten months behind one big, boring rewrite. Claude's new Fable 5 model did it on my first night back. What that week built, and why I credit the specs I wrote by hand.
What went wrong when I built my AI game with Claude Sonnet 4 in summer 2025, pasting code by hand. Six failures, how each one happened, and the rule I still follow because of them.
From June to August 2025 I built an AI monster game with Claude in a browser, pasting code by hand. The systems I had to design myself, and why picking files by hand made me learn my own code.
The first three things I built with Claude in May 2025, coming from Excel and SQL. The model wrote the code, and what stopped me was everything around it, like where files go, the terminal and git.
I gave Claude a vague, misspelled message about my first Minecraft account. It found the skin, worked out why I couldn't log in, and rebuilt a server effect I could only describe by its shape.
Overnight, I had Claude rebuild a strategy game that shut down in 2018, with critic agents judging each piece blind against old screenshots. Where to play it, what it cost and why I'd pick the models myself.
I had Claude play my AI monster-catching game start to finish, fix the bugs it hit, and narrate a presentation from screenshots it took while playing. Why that beat a written report for me.
I had Claude build a skateboarding wizard game from one long prompt in an evening, then ran it again asking for six districts instead of two. How big a one-prompt game can get, and what broke.
I had Claude build a zombie game where you're cave diving. What it would and wouldn't do on its own, why I had it build me a level editor, and who should write a game's song lyrics.
My first post. Why I'm putting what I learn about AI here as short public notes, instead of in scratch files nobody can find, me included.
New entries show up in any reader app — no account, no algorithm, nothing recommended to you. Paste this into one:
aaronorelup.com/rss.xml