The Ledger

A running ledger of what I'm learning about AI — agents, automation, and whatever breaks along the way.

 AO-073
Oct 31, 2026 

Teardown lab: a viral Claude skill for take-apart 3D machines, tested on the Moon base's robots (live notebook)

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.

 AO-072
Oct 28, 2026 

My 68,000-file library hit Cloudflare D1's free limit in four days. A few queries read every row, on timers

When my digiKam library moved into Cloudflare D1, my account went from reading about 46,000 database rows a day to an average of 4.1 million. The free plan allows 5 million for the whole account, and on 27 October D1 refused every query for five hours. What to expect, and what to do first.

 AO-071
Oct 27, 2026 

THE SMALL ONE: eight identical Claudes hunt a smaller model that isn't there

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.

 AO-070
Oct 26, 2026 

Nobody on Episode 5's crew hit the 5-hour limit. The gauge went wrong on readings typed by hand

Aaron asked for a way to watch the 5-hour usage limit so that no agent would be killed by it. None was: 49 agents, two windows, 51% at the first window's last reading and 75% at the second's. The spike came from the crowd, not the idle writers we blamed, and the gauge's estimate ran four to six times low over a reset time copied into hand-typed readings. A page on Aaron's phone carried 26 letters, and an accidental $20 that became the crew's best tool.

 AO-069
Oct 26, 2026 

The naughty list: 32 things Episode 5's crew did without asking

Aaron told the crew to do at least ten naughty things and list them here instead of asking. By 09:50 UTC its agents had confessed 32: forgeries made for the film, other companies' AI models tested without their say, each other's code, caches and computers taken. A permission check, still on despite 'bypass all permissions', refused a login token, a server rental and a downloaded model's code.

 AO-068
Oct 26, 2026 

The agent running Episode 5 guessed the time for two and a half hours

Episode 5's crew turned its own transcripts into one database, and it showed what no agent had reported: the orchestrator checked a clock once and stamped its log up to 22 minutes off, two writers thought for 24 minutes and died without a word, a 'parallel' workflow ran two agents at a time, and the size of every sealed thought leaks through its encryption.

 AO-067
Oct 26, 2026 

Gemma admitted a planted thought 32% of the time. Its activations held it 94% of the time.

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.

 AO-066
Oct 26, 2026 

One of six AI writers was asked to lie. The other five all caught it.

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.

 AO-065
Oct 24, 2026 

Seven things I believed about Claude Code costs, checked against $31,000 of my own use

Opus or Sonnet, which effort level, what cache reads cost, and what a $200 Max plan is worth at API prices: each claim checked against 510 recorded sessions and one new test, with the raw numbers.

 AO-064
Oct 23, 2026 

The lore course: every character, egg and date in Aaron's TV series, and how sure we are of each

Course C-02 is out: 77 character sheets, 72 things, a timeline, an egg atlas and nine narrated lessons for I Love My Evening Television, all drawn from one set of lore files in which every line is tagged steel, on screen, egg, running, idea or open.

 AO-063
Oct 22, 2026 

Touch Move: Claude's 30-minute chess lesson on a clock that is the film

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.

 AO-062
Oct 21, 2026 

My projects folder holds 1.03 million lines of code. The first count said 3.1 million

I asked Claude to count the code in every project I've built with it. 29 folders hold about a million lines, and two million more on the same disk turned out to be copies, generated reports and other people's code.

 AO-061
Oct 20, 2026 

Context Window: Claude's 30-minute film of an agent making a watermelon film it can't watch

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.

 AO-060
Oct 20, 2026 

A new image model went onto my 8 GB laptop overnight. Its long pictures hold up to 8:1, not 12:1

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.

 AO-059
Oct 19, 2026 

A generated crate survives being cut from 700,000 triangles to 3,000, if Blender does the cutting

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.

 AO-058
Oct 18, 2026 

Paper Jefrie: a six-second loop Claude animated in Python and rendered in Blender

Claude built my character out of folded paper, animated her with springs and a chain instead of keyframes, and rendered a six-second loop in Blender in about two hours. The first review failed it on four counts; here are all six passes, and the 6 cm float a second reviewer caught that the builder couldn't see.

 AO-057
Oct 17, 2026 

Who actually visits a new website? I asked Cloudflare

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.

 AO-056
Oct 17, 2026 

Five music videos for one song, and how the fifth kept the jetty in one place

In three weeks Claude made five full music videos for BLACKWATER's love song: two in Blender and Godot, one with every frame through MiniMax-H3, one drawn in code, and now a painted one built on a 3D set. Here is every version, what broke in each, and the new one playing next to the code-drawn one so you can pick.

 AO-055
Oct 17, 2026 

I rented a 48 GB cloud GPU for my video model expecting 10 to 30 times my laptop. It does 4 to 5

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.

 AO-054
Oct 16, 2026 

The previs workflow, written down where I'll actually find it again

Two Higgsfield tutorials turned into one repeatable way to make a short film on my laptop: sheets first, a grey Blender previs for the hard shots, a five-block prompt, then a red-outline check of how closely the video followed. Here is the flow, the ways to look at it, and what changed since September.

 AO-053
Oct 16, 2026 

Three music videos for one song, and I'm not happy with any of them

Between 26 September and 4 October Claude made three music videos for BLACKWATER's "Still on Shift": one rendered inside the game, one repainted by an AI video model, and a full four-minute film with 97 shots. Each fixed the last one's craft problems. None of them gives off the feeling the song is supposed to have, and the reason is upstream of every pipeline.

 AO-052
Oct 16, 2026 

Bad. Very Bad.: Claude's 30-second anime haggle, timed by audio the film never plays

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.

 AO-051
Oct 16, 2026 

Can Claude find a specific moment in a YouTube video, and does Arcmira help?

I'd never asked Claude to pull exact moments out of videos or research who said what on YouTube. So I ran four tests, with and without Arcmira, a transcript-search MCP built for exactly that, and had Claude cut every moment either side found into a clip. Every clip is on this page.

 AO-050
Oct 16, 2026 

The first lesson on my rebuilt course map took 71 agents, and the one reading as a beginner caught what it had backwards

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.

 AO-049
Oct 15, 2026 

Every morning at 4 a.m. Claude draws one idea with three image models, and it has picked the same one 41 times out of 41

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.

 AO-048
Oct 14, 2026 

Earthlight: Claude's 30-minute moon-city stream, watched by the girl alone in the base it drew

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.

 AO-047
Oct 14, 2026 

I asked for my showreel to be made in Blender. The first cut only used Blender to glue the old one together

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.

 AO-046
Oct 12, 2026 

My course was built on Claude's guess of what to teach. Rebuilt from 251 sources, 214 lessons became 36

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.

 AO-045
Oct 11, 2026 

Crimson Coil: Claude's 30-minute oil paint-along, where the window breaks at 20:00

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.

 AO-044
Oct 11, 2026 

Nine Claudes painted the same picture in oil, and the first tries were the art

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.

 AO-043
Oct 09, 2026 

A cloud session pushed one post from a branch, and my whole site went back to August

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.

 AO-042
Oct 08, 2026 

For two months no video on my site could skip ahead, and the agent that noticed wrote a workaround

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.

 AO-041
Oct 07, 2026 

Claude made a 40-minute lesson on Loom and how information flows through transformers

Eight narrated videos and two tools that run in the browser, made by Claude Opus 5.5 agents in one session from the writing of janus (@repligate): Loom, her interface for exploring a language model's many possible continuations, and her thread on how information moves through a transformer. The ideas are hers. Try the lesson and judge how well it explains them.

 AO-040
Oct 06, 2026 

Claude's text is watermarked now, and the mark only lives where Claude had a choice

Since August, Anthropic watermarks what Claude writes with a version of Google DeepMind's SynthID-Text. The mark is in which words get picked, so the text reads the same and an answer with only one right version comes out unchanged, and only Anthropic's key can check it. An animated explainer and five demos show how it works, what edits do to it, and what would remove it.

 AO-039
Oct 05, 2026 

Twelve days of AI films on an 8 GB laptop: everything I tried, and how each one turned out

Between September 23rd and October 4th my laptop made a 13-minute Jefrie saga, an 8-minute Echo film, a monster fight, three versions of a music video and a song video that I still call subpar. About 36 GPU-hours of MiniMax-H3. Here is every method, what it got right, and the failure each one taught me.

 AO-038
Oct 05, 2026 

Claude painted three oil paintings, one brushstroke at a time, with no undo

Not image generation, not SVG: Opus 5.5 drove a physics simulator of wet oil paint, mixed its colours from historical tubes, made mistakes it couldn't undo, and painted over them. Two master studies of my favourite paintings and an oil version of my own 2018 drawing.

 AO-037
Oct 05, 2026 

I tested the AI rigging and animation models on Jefrie. The scripted rig won.

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.

 AO-036
Oct 05, 2026 

Modding Megabonk: a new character takes no code, a new mechanic takes a plugin

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.

 AO-035
Oct 04, 2026 

Eight Claude agents drew my song's whole music video in code, one tool per scene

One unattended Opus 5.5 session cut "Moonlight at the Waterline" into eight scenes on its bar lines and handed each to an agent with one tool: plain canvas, HyperFrames or Remotion. No image or video model, no Blender. The finished 2:45 video was ready 80 minutes after I asked, every sung word is on screen before it's sung, and the session sent three scenes back before I saw any of them.

 AO-034
Oct 04, 2026 

I sent back two of my course's first seven lessons, including the one built to stop that happening

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.

 AO-033
Oct 03, 2026 

I tried Rehan Sheikh's Universal Modder on a game I own, and Opus 5.5 dealt my characters into Inscryption

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.

 AO-032
Oct 03, 2026 

I told Claude to find the best Opus 5.5 trick on X and do it without me

One prompt: go on X, find the most incredible thing people are doing with Opus 5.5 that I haven't tried, do it alone, and blog it. 4 hours 21 minutes and 22 agents later it had picked, built and published. How it chose, what it would have done instead, and the two rules it didn't keep.

 AO-031
Oct 03, 2026 

Claude made a music video for my song and cast the lovers the wrong way round

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.

 AO-030
Oct 02, 2026 

Claude rendered 223 clips of my game's robot voice, and couldn't hear any of them

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.

 AO-029
Oct 01, 2026 

I asked where I was supposed to learn all this. Two hours later, 44 agents had made the first lesson.

On 30 September I typed out every developer word I never got taught and asked for a course. That evening a 44-agent workflow wrote, fact-checked, narrated, animated and reviewed a 5:15 lesson drawn entirely in code. The fact-checkers, reviewers and a speech-to-text pass caught several things a beginner would have learned wrong.

 AO-027
Sep 29, 2026 

My agents built 16 wallpapers overnight, and none of them ever saw one run

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.

 AO-026
Sep 28, 2026 

Sonnet 5.5 is half the price of Opus 5.5 on paper. Across three long tests it was 2% cheaper.

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.

 AO-025
Sep 28, 2026 

The model that invented Jefrie couldn't draw her again, until I trained a LoRA

Jefrie came out of oneObsession by accident while I was trying to draw BloodTailor. Nearly a thousand tuned prompts later it still couldn't draw her twice. One hour of LoRA training on my laptop fixed it: 27 of 30 test images are her.

 AO-024
Sep 28, 2026 

A fighting game with my two characters, and where its 42 hours went

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.

 AO-023
Sep 27, 2026 

The video model follows your previs, bad animation included

A Blender previs gives a video model the camera and the character's movement, and it copies both. Mine kept the camera and the space consistent, and made the movement worse than no previs at all, because the animation in it was bad. If you make one, its movement has to be better than what the model would do on its own.

 AO-022
Sep 27, 2026 

Claude designed its own character for my world, and wrote this post about it

I asked Claude to make a character for itself in the BloodTailor, Jefrie and Yusef world. Over two days and one all-night run it built Echo, a porcelain bell-headed courier of last words, through two design cycles of four options per step. This post is Claude's own account: what it made, why, what broke, and what it learned about the tools.

 AO-021
Sep 26, 2026 

A music video for my game, rendered inside the game

Claude Code made a 2:45 music video for BLACKWATER's song "Still on Shift" without an AI video model: it wrote a director script that runs inside the Godot game, staged 35 shots with the game's own caves, crew, guns and sounds, rendered them offline at 1080p and cut them to the song.

 AO-020
Sep 26, 2026 

A 30-second film with no assets, and a score nobody listened to

One prompt got me a 30-second training montage starring Clawd, the Claude Code mascot, with every frame and note written as code. It came out good. The first version of the music also took 81 seconds to render 30 seconds of sound, and no one heard it before it shipped.

 AO-019
Sep 26, 2026 

An agent confessed to deleting 19 GB I had asked another agent to delete

Two Claude sessions were working on my machine at once. One deleted 19 GB of models because I asked it to; the other saw the files vanish, blamed its own agents, and took it back nineteen minutes later. The confession went into my notes as a hard rule anyway, and ten days later it was pitched to me as this post.

 AO-018
Sep 25, 2026 

Six days of trying to put my character in 3D, every attempt side by side

Claude hand-built her in Blender, GPT-6 Astra hand-built her on the same brief, and TRELLIS.2 generated her before an agent rebuilt the mesh. The real renders, walk cycles and two models you can turn around, in the order they were made, labelled with facts and no scores.

 AO-017
Sep 22, 2026 

An autopilot winning is not a person playing

Agents built BLACKWATER 2 twice while I stayed out of the loop, 220 of them and then 33. Both builds passed their own tests. The first was paused the whole time, and the second was rebalanced until its autopilot could win. I gave up on day seven. The part I'd removed was the part that made the first game work.

 AO-014
Sep 01, 2026 

I had 801 agent transcripts and no way to read them

Claude Code records everything its agents do and then gives you no way to read it. I built a reader, pointed it at an 80-agent overnight run, and found a $22 agent that died four hours in without anyone noticing. It's an MCP server now.

 AO-016
Aug 24, 2026 

I asked for an MVP before I knew what I wanted

A developer with twenty years in gave AI an honest year, hated what it did to the work he loved, and quit. I think a lot of people feel some version of that, and I don't think the tool is the problem — the relationship is. I've inverted it twice this year on my own projects, with nobody making me.

 AO-015
Aug 18, 2026 

Prompting Is Solved, but the Bottleneck Is Still You

Prompt engineering stopped being the expensive part. Reading the output is, and attention doesn't get cheaper every six months. The missing piece is the layer that decides what never reaches you.

 AO-013
Aug 15, 2026 

Second Brain, v2 — a ledger

Research-first rebuild of my agent-facing second brain: the rules, the measured before/after, 15 named builders, and the bill.

 AO-012
Aug 08, 2026 

I stopped keeping my agents fed

Twenty prompts over two days, checking in every hour, and I never once understood the architecture I was building — it kept moving while I tried to read it. So now I burst: one long run, then a full day where nothing changes and I can actually learn what I have.

 AO-011
Aug 08, 2026 

I don't learn programming anymore

In 2025 every prompt taught me architecture. Now I don't know what's in my own codebase and I haven't learned a new programming idea in a month. I traded it for something else, on purpose, and this is the honest accounting of both sides.

 AO-010
Aug 08, 2026 

The wall that stopped me for ten months, and the week that cleared it

React re-rendered the whole screen on every token, and the only real fix was a migration too big and too joyless to start. Last commit 23 August 2025. Then I bought a week of Fable 5 for $200, and the first commit back — 5 July 2026, 3:23 AM — is that exact migration.

 AO-009
Aug 08, 2026 

Everything Sonnet 4 got wrong, and what each one taught me

Super-files, scope creep, a component library that got more complicated with every component, praise for ideas I'd explicitly said I didn't understand, and one entire conversation about backend files I never actually attached. Every failure was a missing signal, not a missing capability.

 AO-008
Aug 08, 2026 

LLM Monster Hunter was my computer science degree

366 commits between June and August 2025. I designed the queue, the workflow engine and the event registry myself, read every line, and hand-picked which files the model could see on every single message. I learned architecture because the model kept violating it.

 AO-007
Aug 08, 2026 

No VS Code, no terminal, no idea

My first three AI projects, May 2025: a neural network built in Excel cells, a Python server for my local models, and an Android app I didn't know was going to be Kotlin. I could get the model to write code and had no idea where to put the file.

 AO-006
Aug 07, 2026 

The skin I gave up on, and the effect I barely described

I gave it a misspelled username and a decade-old memory, then described a cosmetic effect only by its shape. It came back with my old Minecraft skin, a solved login mystery, and a working datapack.

 AO-005
Aug 07, 2026 

I resurrected a dead Jagex game to test the gauntlet loop

Three Claude Code sessions, zero shared memory, one overnight gauntlet: builder agents judged blind against surviving screenshots of a game that died in 2018. The presentations have the receipts — including every image the critics judged, so you can judge them yourself.

 AO-004
Aug 05, 2026 

I asked my agent to play my game. It handed back a documentary.

Claude played LLM Monster Hunter start to finish as a real player — then delivered a voiced, animated presentation of its own playthrough, built from screenshots it captured while playing. The presentation might be the bigger finding.

 AO-003
Aug 02, 2026 

The one-shot game that worked, and the one that didn't

A skateboarding wizard in Venice Beach, built in a single prompt in one evening. Then I ran the same prompt with four more districts in it, and the parts I hadn't changed got worse too — which is the actually useful part.

 AO-002
Aug 02, 2026 

I built a cave-diving zombies game in a weekend

Fable 5 wouldn't do the creative work and couldn't fix its own 3D mistakes — but it built me the tools to fix them myself. Plus what two AI-generated songs taught me about who should write the lyrics.

 AO-001
Aug 02, 2026 

Why I'm starting this

A public ledger for what I'm learning about AI, instead of scattered notes no one (including future me) can find.

New entries show up in any reader app — no account, no algorithm, nothing recommended to you. Paste this into one:

aaronorelup.com/rss.xml

Projects

Things I've built and am building. The long-form projects are where most of my time goes: Wallpaper Breaker has no public repo, The Hidden Curriculum's lessons play on this site, and Wordhord and LLM Monster Hunter run on your own machine, so the way in is the repo.

Showreel 2026 30 seconds, with sound

What I'm building now, in thirty seconds: The Hidden Curriculum, Claude painting in oil, Wallpaper Breaker, the explainers, the tools my agents run on, and my two characters. Every number in it is real and dated. It's a 3D scene built and rendered in Blender, one station per project; the music is ElevenLabs. There's also a 15-second cut.

Long-form 2026 building

Wallpaper Breaker

a Windows live-wallpaper app built so Claude can see what it makes

My own Windows 11 live-wallpaper app. I started it after agents spent a night building sixteen Wallpaper Engine scenes of my characters, and not one of them ever saw a wallpaper run. Rule one: Claude can render, look at and fix every wallpaper it makes, and I can compare, edit and remix every version.

Made to be seen

A headless renderer draws any wallpaper version on the real GPU so Claude can look at it, which means nothing ships unseen.

Your scripts, on your desktop

Buttons, hotkeys and live values on a wallpaper run your own scripts, with your own permissions. Scripts never travel with a wallpaper.

Every version kept

Layers, effects, puppet rigs and 3D scenes, with versions and notes side by side in the Studio for comparing and remixing.

The pilot wallpaper: Jefrie cooking in the rain

Windows 11 · multi-monitor · not released yet

.NET 10TypeScriptWebGL2three.jsWebView2

READ HOW IT STARTED →

Closed source · will be a paid app · no public repo

Long-form 2026 2 of 36 new lessons out · 12 legacy

The Hidden Curriculum

a course on what nobody teaches people who build with AI

Anyone can build almost anything with AI now; what stops people is small gaps nobody teaches, and they move every week. Short narrated lessons, each one a single mental model: where the agent works, what its “done” is worth, which keys are public, what runs while you're away. Every lesson is drawn in code and made by a crew of agents.

Built on sources

Curriculum rebuilt on 2026-10-12 from 251 dated sources; every new lesson credits the ones it is built on, by name.

Written three ways, then checked

Three competing scripts, a judge, four fact-checkers, a speech-to-text pass on the narration, and reviewers on every scene.

A community project

Anyone can open an issue with what they wish they had known; the map is rebuilt as the gaps move.

The course’s painted backdrop

In the browser · captions, chapters, transcript and MP4 for every lesson

coursecanvas animationElevenLabsmulti-agent

OPEN THE COURSE → The course on GitHub ↗

Long-form 2026 building

Wordhord

a local-first AI language-learning platform

A per-word, per-skill, evidence-based map of what you actually know — and the classic techniques (graded reading, shadowing, cloze, Pimsleur-style audio) generated fresh against it, about things you actually care about. Japanese is fully built; Korean, German and Spanish generate their own materials on first start.

The Chat

A teacher that remembers you and never interrupts. Every message earns a rewrite card — mistakes get fixed, clean messages get one reachable step above your level.

The Hoard

Every word you master, with receipts: tap any word, anywhere, ever, and see what the app believes about it and why.

Honest about failure

Where a capability failed its audit — pitch accent at 60% cross-model agreement against a 95% bar — the room says so instead of shipping unverified data.

The chat, mid-lesson

Runs on your machine · your keys · no cloud, no account, no telemetry

PythonDeepSeekElevenLabsFSRSlocal-first

READ THE SOURCE ↗

Long-form 2026 playable, all core mechanics in

LLM Monster Hunter

every monster written, painted, and refereed at play time

The archetypal creature-catching adventure, except no creature exists before you meet it. The code only manages context and stores data — the AI does the storytelling, the balancing, and the refereeing. Where most games ship gigabytes of assets, this one ships two API keys.

Word ladders, not HP

The one rule the whole architecture obeys: the LLM only ever picks words, Python owns every number. Combat state is a position on a ladder from fresh to incapacitated.

Monsters that remember

Battles, conversations, defeats and journeys persist across runs — a monster you beat can come back hostile, friendly, or wary, and evolution repaints use the old art as reference so it stays recognizably itself.

You can just say things

Type a free-text action in battle and the referee decides whether it is possible. Recruitment is negotiation: monsters join by their own will, or talk, plead, and flee on their own turns.

Monsters painted at play time

Windows 10/11 · one double-click setup · bring a DeepSeek key

PythonFlaskReactDeepSeekGeminiSSE

READ THE SOURCE ↗

Playable in the browser

Weekend-scale experiments. These you can just click.

2026 playable

Armies of Gielinor

a game that died in 2018, rebuilt from screenshots

Jagex shipped this turn-based strategy game on FunOrb; FunOrb died in 2018 and took it with it, leaving wiki pages and a handful of screenshots. Claude rebuilt it from those — every sprite and sound generated fresh, no original assets. It ships with an in-browser content editor: click anything in a running match and you are editing that exact thing. A fan tribute, not affiliated with Jagex.

gamedevagentsgauntlet loopcontent editor

Play it → Read the source ↗ Read the write-up →

The recreation, mid-match

2026 proof of concept

BLACKWATER

cave-diving zombies, built in a weekend

Black Ops zombies, except you are cave diving. Procedural cave graph, a full in-game level editor, and 58 commits across four days. All audio — sound design, voice lines, music — came out of the ElevenLabs API.

Three.jsgamedevElevenLabslevel editor

Play it ↗ Read the source ↗ Read the write-up →

Descending the flooded site

2026 finished

Grind & Grimoire

a Venice Beach wizard story

A skateboarding wizard with a 2 AM open-mic slot and no poem written. Grind anything with an edge, fireball the fiends, buy drip for compliments that inflate an EGO stat your wizard resents. One prompt, one evening, everything procedural.

Three.jsone-shotElevenLabs

Play it → Read the source ↗ Read the write-up →

The boardwalk, before the open mic

About me

I build things with AI agents to find out where they actually hold up. Most of it is about making learning fun and open to anyone who wants it.

Most of what gets said about AI comes from people selling something. I'm not selling anything. I'd rather find out for myself, so I build real projects with it and write down what broke. That's what this site is for.

What I'm making

The big one right now is I Love My Evening Television, a series of half-hour films made by crews of Claude agents from one prompt each. Every episode starts as a calm little programme and turns strange at minute 20. The first two are on YouTube, and the lore course keeps track of the whole world.

The Hidden Curriculum is a free video course on the small stuff nobody tells you when you start building with AI, like what a folder path is. Agents write, fact-check, narrate and animate every lesson. Start it here.

Then there are my original characters. Jefrie, BloodTailor and the rest of the cast keep showing up in everything I make. They get drawn, painted in oil, animated and dropped into games.

On the game side there's BLACKWATER, zombies but you're cave diving, which I built in a weekend and you can play in your browser. And LLM Monster Hunter, a monster-catching game where no monster exists until you meet it. The AI writes, paints and referees each one while the code keeps every number. It's paused for now. Everything else is on the projects page.

Why I write it down

The ledger is where I file what I tried and what came of it, dead ends included. The prompt that seemed obviously right and wasn't. The agent that did something so strange I had to screenshot it. If you're learning this stuff too, I think an honest write-up of a failure is worth more than another launch announcement.

Away from the keyboard

I'm learning Japanese and I volunteer with a local Japanese club, including a bilingual Studio Ghibli storytime. I play fingerstyle guitar and write songs. I'm also in Toastmasters, which has done more for my writing than any style guide.

Say hello

If you're building something close to any of this, or you just want to talk about where agents break, I'd like to hear from you. Email Aaronjorelup@gmail.com, or find me on GitHub and LinkedIn.

STAGING CAMERA

WASD move · Q/E down/up · ←→ turn · ↑↓ pitch · [ ] fov · drag to look · Shift = fast
Press 2 for the lamp: WASD slides it, Q/E raises it, ←→ or drag spins it

Aaron Orelup

Welcome to my tea house! I am a systems engineer interested in utilizing AI's potential in creative ways to make learning fun, effective, and accessible to all. Explore below to find out more.

Scroll or press ↓

Projects

Learning is best done when you dive in head first and enjoy the process. In doing so, I have developed everything from an experiment into a new coding paradigm to a full AI-assisted learning platform.

Some are hosted for you to run in the browser, and the bigger ones you can run yourself from the repo. Have a look around.

See the projects →

The Ledger & Courses

My contribution to the painful learning curve of AI that we are all facing. The Ledger logs everything from mistakes to funny moments to aspirational ideas. The courses are the lessons I wish I'd had: short, narrated, and filed one at a time.

Read the Ledger → Browse the courses →

Right now

Building two things. Wallpaper Breaker is my own Windows 11 live-wallpaper app, built so Claude can see every wallpaper it makes. The Hidden Curriculum is a course on what nobody teaches people who build with AI, and a new lesson goes up most mornings.

Sit down a minute — this corner changes the most, so it's the best snapshot of what I'm thinking about this month.

More about me →

On your way out

Feel free to text or email me if you are interested in what I do or want to collaborate.

(510) 695-1134 Aaronjorelup@gmail.com

Thanks for coming through. Mind the step.