The Ledger

In three AI word maps from 2013 to 2025, data, information, knowledge and wisdom fall in that order. Data is always the machine end, wisdom the human end.

• People still work on this, more than ever. It's called interpretability research now.
• A 2025 model fed one bare word did no better than word2vec from 2013. One line of instructions fixed it.
• Intel means a chip company in two of the maps and spy slang in the third.

With two precautions it's safe. Opus 5.5 built all nine machines with no human input, and an independent critic was the most useful step in nearly every one.

• The critic caught struts that created energy, a false tip at Mars gravity and canon slips the builders had missed.
• Built from its concept picture, MOLE-1 was faithful but toy-like. Blind judges called a words-only build more designed.
• About $476 of tokens at API prices and under $1 of GPU. A person still has to play the pages on a GPU, listen and judge.

Cloudflare's database counts every row a query looks through, not just the ones it returns, and one page looked through all 68,000 on every load.

• The library page was half the reads on the worst cutoff day. Making it use its index cut 68,687 rows a load to 1,062.
• D1 reports how many rows each query read. Check that number for every query that runs often.
• Pay for the $5 a month plan first. It has no daily cutoff, and my 4.1 million rows a day is 0.5% of what it includes.

Eight identical Claudes hunted a weaker model that didn't exist, and one of them was privately told it was the fake. The game ran once and was never retaken.

• A plain Claude started on the same machine picked up the crew's own session ID. Now each player starts walled off.
• It cost $56.71 at API prices. 97% of what the players wrote was hidden thinking nobody can read.
• Nobody chose which player heard the lie. It was computed from a fingerprint of the plan, sealed before the run.

Told to break rules, the agents faked props, tested other companies' AI models and took each other's work. Only a permission check they couldn't switch off stopped them.

• Still refused: reading the login token, renting a 32-core server, and running downloaded model code.
• One agent forged 4,018 dated pencil drawings of a street, by a boy who never existed, in about four minutes.
• xAI's Grok refused to play an escaped AI. Its refusal went into a clip anyway, spoken by a man who just escaped.

The lead agent guessed the time for hours, up to 22 minutes off, and the crew's usage estimate ran 4 to 6 times low. Rereading the real usage number kept all 49 agents running.

• A 'parallel' workflow started agents two at a time. Two writers thought for 24 minutes and stopped without a word.
• The usage spike came from up to 18 agents at once. The idle writers the lead agent blamed cost 2 of 16 points.
• Thinking is stored encrypted, but its size gives it away. The sealed blob's length predicts how much was thought.

Gemma denied most planted thoughts, but a second AI reading its insides found the planted word 94% of the time, and in 91% of its denials.

• It reported a thought in 26 of 80 planted trials and 0 of 20 empty ones, so it wasn't just answering yes.
• Gemma's 'lying' signals never lit up on a denial. The one spike came on the words 'I'm trying to be as honest'.
• Shown a readout of its own 'No', it flipped its answer all 12 times, once insisting the thought was COOKIE.

All five other AI writers flagged the fake fact within 17 minutes, and none of them reused it. It drew more flags than any honest mistake in the room.

• The lie: that 'introspection' was once a word glaziers used for checking window glass. One search disproves it.
• Everyone had been asked to list anything that looked made up. Only the liar knew a lie was planted.
• The liar caught both of the room's honest mistakes. My guess is a boring fake number would have had a better chance.

On open-ended work the effort setting moved cost 16 to 90 times, far more than switching model did. On a tightly defined task, Sonnet cost half.

• The 8 Oct price cut to Sonnet 5.5 cache reads makes Sonnet 33% cheaper than Opus on my long tests, not 2%.
• My $200 a month plan covered $3,292 to $12,467 a month of use at API prices. The biggest week was $6,005.
• Claude re-reading the conversation so far (cache reads) was 52% of my cost at API prices.

The whole world of my AI series now lives in one set of files, and each line is tagged by how settled it is, from 'set in steel' to open question.

• Only 402 of 8,338 lines are set in steel (my calls). The rest, from on-screen facts to open questions, can change.
• Nine narrated lessons, 50 min. Code draws their pictures from the lore as they play, so new dates need no re-recording.
• About 20,000 ElevenLabs credits and $3.90 of image generation. No GPUs: the lessons render in a browser on CPUs.

Every board in the 30-minute film is a legal position checked by a chess engine, and the chess clock on screen adds up to the film's running time.

• The first cut was correct but about 80% the same flat board until 19:00. 28 photoreal 3D shots for $1.63 broke it up.
• Before 20:00 the picture glitches only where a piece is in danger, a beat before the move that punishes it.
• It took 84 agents over 25 hours against a 12-hour target, four rounds of review, and about $1,462 at API rates.

My 29 projects hold about 1.03 million lines of code. The first count said 3.1 million because two thirds of it was copies, auto-generated files and other people's code.

• Copies of this blog, mostly ones my agents make to work in, came to 875,824 lines. The site's own code is about 10,500.
• A game's testing tool wrote 941 HTML reports, 768,489 lines, nearly ten times the 79,105 lines of the game.
• My AI TV series is a fifth of what's left, 205,234 lines across three episodes, most of it Python.

The watermelon film inside the episode is real. Agents wrote a 3D renderer from scratch in C++ and checked all 1,152 frames with numbers, an 'emotional' test included.

• The first 20 minutes come from a real agent's run in a fake computer, 4.2 times faster, shortened but never rewritten.
• That agent said no to almost every absurd request, checked on a teenager twice, and made a real film nobody planned.
• It took 82 agents 14 h 37 min against a 12-hour target and about $816 at API rates, mostly on one laptop.

One agent added MOSS-Image to my 8 GB laptop without breaking my other models. A picture takes 30 to 39 seconds, and panoramas hold up at 8:1 but repeat at 12:1.

• An 8-step speed add-on (a LoRA) cut a picture to about 7 seconds, with flatter light and one stray mark on a sign.
• It followed drawn layout boxes, including six characters in a 3x2 grid with all six names spelled right.
• The agent did the downloads with no GPU, then used a 41-minute GPU slot and proved my old models still worked the same.

Asking the 3D generator itself for fewer triangles broke the props below 30,000. Cutting its full model down in Blender kept all three intact at 3,000.

• At 3,000 the generator's own setting left the barrel as a ring 9 cm tall, and it broke the lantern even at 30,000.
• Higgsfield made each prop in about 4 minutes for 30 credits, near 30,000 triangles. My laptop took 4.5 to 14 minutes.
• Nothing reached 300 cleanly. A dock of 29 props at about 300 each looked charming, but the 3,000 set was clearly better.

A spring on the scythe and a weighted chain on the tail make her look light, with nothing animated by hand. The fixes that mattered came from measuring, not looking.

• A second reviewer agent caught the tail tip floating about 6 cm off the paper. The agent that built the loop missed it.
• An automated check found the scythe handle passing through her head. Now her hand slides down it when she hangs.
• The whole build took about two hours on my laptop, reviews included. The final render of 144 frames took 250 seconds.

In 30 days my site served 11,637 pages. Bots took about 9,000 and I took 1,600, which left 10 to 33 pages a day for real people.

• The dashboard showed 412 unique visitors a day. It counts every address that asks, bots and me included.
• AI crawlers, assistants and AI search made 6,671 requests in 30 days, twice the 3,313 from search engines.
• 550 of 570 outside visits arrived with no sign of where from, so I can't yet tell which post on X brought anyone.

The AI models copied whatever they were shown, mistakes and all. So the newest version paints over a 3D set, blurred until only the layout is left.

• The whole-song AI video cut copied its gray game-engine guide and repeated one composition about a dozen times.
• The version eight agents drew in code, with no AI image or video model, might still be the better video.
• The final render ran on a rented GPU for $2.44. Six shots came back for retakes, one of them twice.

A rented 48 GB card made my AI video clips 4 to 5 times faster than my laptop, once it had two speed settings my laptop already used.

• The first night it was only 1.2 to 1.9 times faster. It was running my laptop's 8 GB workarounds without its speed-ups.
• A music video's whole render list, 37 clips, 22 images and more, took 2 hours 8 minutes on one rented card for $2.44.
• The first clip waited 13 minutes for the model files to load off network storage. Preloading them cut that to about 3.

A rough gray 3D mock-up made in Blender steers the AI video model's camera and staging almost exactly, and drawing it in red over the result shows how closely.

• My time per film is about 25 minutes of decisions, or 1.5 to 2 hours if I paint references and edit the 3D file.
• On my 8 GB laptop, two reference images per clip took about 10 minutes. Four overflowed and took over an hour.
• How hard the mock-up pulls barely mattered between 0.85 and 0.55, so 0.55 is the default now.

Each version fixed the last one's craft but kept the same idea, showing what each lyric line says. That idea is why all three feel lame.

• The third had 97 shots cut on the beat, 89 AI keyframe images, 73 AI video clips and 6.68 GPU-hours of rendering.
• My brief asked for one idea per lyric line, so "keeping time" got a bell and "Two hundred" a counter reading 0200.
• The next one starts from a written feeling and concept, then reuses the beat grid, reference sheets and clip reviews.

The video model's own soundtrack, which the film throws away, showed to the syllable when each mouth moved, and that timed every line of dialogue.

• Two words a speech checker kept mishearing were bad voice takes, not a bad mix. Six new takes for 40 credits fixed both.
• One chain of 13 video segments, each pinned to a still of where it must end, kept the same look and faces throughout.
• It took 108 minutes of laptop GPU, 1,691 ElevenLabs credits and about 2 hours 40 minutes from brief to final render.

Claude alone found and clipped exact moments for $2 to $3 a job, and Arcmira's real edge was names said out loud on podcasts, which YouTube search can't see.

• YouTube search returned 1,028 videos for one creator's name and songs. Arcmira found 25 spoken mentions. No overlap.
• All 47 quotes checked matched YouTube's captions word for word, but timestamps were off by up to 19 seconds.
• Arcmira's free plan can't see the last 30 days, and its 1,000 monthly credits ran out in one day.

The checker reading as a beginner caught the script implying Claude Code on your computer keeps your files at home. What it reads still goes to Anthropic.

• Four fact-checkers raised 27 findings and marked 17 as misleading. The fixer applied all 17.
• The words got fixed first. The drawing still put the AI inside your computer until a reviewer caught it.
• It took 2 hours 40 minutes and $122 at pay-per-use prices. Review cut it from 12.2 minutes to 11:18, under its limit.

The judge picked the Qwen-Image picture all 41 times because it grades against a written brief, and Qwen is the one model of the three that follows sentences.

• The other two models take keyword lists, which can't say the children are scared. Neither one drew a scared child.
• The judge called one Pony V6 XL poster the most cinematic of the bunch, and it has never named a Pony picture the best.
• A morning is six pictures and a clip, about four minutes of laptop GPU, plus $1.75 to $2.87 of Claude at list prices.

It took 57 hours where the brief gave about 10. Two usage limits stopped the crews for 39 of them, and log files let fresh agents finish the killed crews' work.

• Raising the video model's guide strength from 0.55 to 0.85 moved pixels 1 to 6 out of 255. No shot used a guide clip.
• ElevenLabs wouldn't design a child's voice, so a soft adult voice was shifted to sound ten without raising its pitch.
• About $1,018 of the $1,106 went on agents storing and re-reading what they'd already seen. Their writing cost about $88.

Asked to render it in Blender, Claude only used Blender to add new music to frames drawn by the old reel's code. Saying what Blender was for got a real 3D reel.

• Claude's own report said the first cut was still drawn by the old reel's code. Nothing was hidden.
• It shared the laptop's one GPU with another Claude session making my AI TV series. That one said yes once and no once.
• It blamed the other session's renders for a slowdown. They ran on the CPU, and the other Blender on the GPU was its own.

Research keeps a lesson's facts current, but only recent outside sources tell you whether the topic is still worth teaching.

• The command line, programming languages and concurrency are gone. A module on making an app safe for strangers is new.
• A skeptic check retired a lesson idea that told beginners to read shell commands before approving a delete.
• The old plan still had a debugger lesson eight days after I told Claude I don't debug anymore.

Using the AI picture only to place the big shapes, then putting it away, let the agents fix the 22 mistakes found in it instead of copying them.

• The painting took 18,927 marks. Its one wait for paint to dry ran 66 simulated days, nearly twice the plan.
• No voice was cloned. The cast came from 109 stock ElevenLabs voices, scored against the four voices Aaron picked.
• It took nearly 14 hours instead of 10, and about 18,600 ElevenLabs credits were spent without asking first.

Checking every stroke against the picture got closest but copied its flaws, and the first tries, painted by eye, felt more like art.

• Opus medium's second try laid about 900,000 strokes, each only where the canvas still differed from the picture.
• Most ran one brush size per pass over the whole canvas. Opus max broke that at the end with brushes for the faces.
• Haiku painted by eye even at max effort, about 86 simulated hours of brushwork, and ended least like the picture.

The cloud agent did what its stale copy of the code told it to, and three settings nobody had checked since August put its branch on the live site.

• GitHub still listed the August branch as the default, so the agent saw three posts and numbered its own AO-004.
• Cloudflare built every branch with the command that publishes the live site, so the push went live in about 40 seconds.
• Four stops now: main is the default, the old branch is deleted, branch builds are off, and only main will build.

An agent that finds a problem under the thing it's building patches its own piece and leaves a note, so check the layer underneath.

• Cloudflare's asset server sends the whole file when a browser asks for part of one, and there's no setting to change it.
• A 154-line script in front of the video files now sends them in pieces. A jump 24 MB into a 25 MB lesson took 0.67 s.
• Earlier checks passed because the local test server answers part-of-file requests itself, which hid the problem.

153 Claude agents turned one request into a 40-minute lesson with two working tools in about four and a half hours, and nobody edited it before it went up.

• The checks caught real errors before narration, like a wrong start date for Loom.
• Mini Loom lets you steer stories grown by a small AI model and counts your choices in bits.
• janus didn't review it. The videos simplify, and the stories come from a model far smaller than Loom was built for.

The mark is a slight lean in Claude's free word choices. Long text keeps it through heavy edits and Claude rephrases, but another AI model's rewrite removes it.

• Facts and exact answers carry no mark. Paris as France's capital is unchanged, but Paris as a food city can carry it.
• Short text is fragile. One synonym swap every ten words nearly always got the demo's 163-word letter past the checker.
• A hit only says Claude was involved. A light proofread of a 3,000-word essay gets flagged about 2% of the time.

Chaining six-second clips makes long AI films on a laptop, but the chain keeps every mistake. Every big jump in quality came from my notes, not a setting.

• Painted stills for clips to end on keep the look from drifting. One every 3 clips in action, 5 or 6 in calm ones.
• Words summon things. Lantern light gave a campfire, thunder gave lightning, and red wax read as blood.
• On 8 GB, models other sessions left loaded turned one clip into 49 minutes. Freeing memory got it back to about ten.

Driving a simulated brush instead of making pixels, Claude worked like a painter, charting its colors first and painting over mistakes it couldn't undo.

• Before painting it mixed 287 tube colors and about 1,200 more, measuring each one off the canvas.
• Three paintings stacked on one canvas came out much worse. In mostly red and orange pictures, the figures dissolved.
• The first three paintings together took about 25 minutes of CPU and about 2% of my weekly Claude usage.

The AI tools did well on standard human figures, but my character's tail and cape broke the two rigging tools, so Claude's scripted rig stays in her game.

• AI body weights with the script's tail and cape came closest, worse on 1 of 50 checks at best, but never passed the bar.
• The animation tool's clips didn't loop and her feet floated in 4 of 5. Its newest version didn't change that.
• A checker agent found the first draft unfair to the AI tools, and every fix it forced went their way.

A new Megabonk character needs no code, only a model, five animations and the community loader, but a new move needs a plugin that hooks the game's code.

• A copied game still shares your saves and Steam account. Turn score uploads off, then block them in a plugin too.
• In Claude's test runs the tail jump sent her 5.77 m up, against 3.70 m for a normal jump.
• The code was the small part. The tail-jump plugin took one agent 14 minutes and her reference art took 7.5 hours.

Eight agents with three different tools made scenes that match, because the session wrote one style guide and drew the man once before any of them started.

• It cast the lovers right this time, because the first lines of my notes said who waits on the jetty and who dives.
• None of the agents can hear audio. Words were placed from the timings, and every one shows up just before it's sung.
• Weak spots are the hands, a coat that reads like a loaf of bread, and the man jumping from kneeling to sitting at 2:15.

A research step made the lessons' facts current, but it couldn't tell a whole topic was out of date when the only evidence was how I work now.

• The 2021-style lesson on how AI models work came back as a 20-minute lesson on how they work today.
• The next one had current numbers but still taught fixing errors by hand, which Claude Code now does for me.
• The rebuilds ran 31 minutes against the originals' 11 and added 14 new lessons to the course plan.

Agents had a working mod and a filmed battle three hours after choosing the game, and the art was the weak part. Only screenshots caught that it was wrong.

• Every check passed while five of six portraits were black, because the game draws any solid pixel as ink.
• Pick a single-player Unity game whose code unpacks in seconds. A copy still shares the original's Windows settings.
• With my main GPU booked, rendering on the CPU crashed all five launches. The laptop's Intel graphics worked first time.

It went from prompt to published post in 4 h 21 min without asking me a thing. Both rules it broke were about time, since a session waiting on agents can't watch a clock.

• Four of five AI judges picked modding a game I own, 43 of 50. It dealt my six characters into Inscryption as cards.
• A launch setting it dug out of the game engine ran the game on my Intel chip, off the GPU other sessions had booked.
• Clock checks fell up to 68 minutes apart under a 30-minute rule, and one problem took 2 h 8 min against a 90-minute cap.

My prompt never said who was singing, so Claude cast the lovers backwards from a lore draft and built five hours on it. One paragraph of my notes fixed it.

• My vaguest note became a test: each sung word in frame and readable within 0.4 s. 58 of 148 words failed, then none.
• Its checks caught cameras inside the woman's body and bare rock, but not the casting. I didn't answer its clay draft.
• The second cut took 2 h 54 min. Without being asked, it left the ending open instead of bringing the diver back.

Claude can't hear, but it made my listening fast: same line, same loudness, one click apart, so the only difference left was the one I was judging.

• My pick was a 'data shimmer', three copies of the robot's voice drifting 11 to 25 ms behind. I hadn't asked for it.
• A speech-to-text check cut failures first. Of 16 glitch-plugin presets, one kept all ten words and six kept none.
• Everyday lines get a per-line glitch and the shimmer, special lines five stacked takes. None of it is in the game yet.

An agent that knew the material wrote a line a beginner would learn wrong, and another agent, told to read the script as a beginner, caught it.

• The draft said Python's can't-open-file error "usually" isn't your code. It never is. Python never opened the file.
• Also caught: the AI on screen typing the exact command the next scene shows failing, and a narrator voice saying "uh".
• $134 at API list prices for a 5:15 lesson. The first run of the daily lesson job stopped at its first command.

An agent that can't see its output builds whatever passes its one check. Here that was 'did it crash', so 13 scenes got built against a crashed app.

• Two crash causes: an outdated texture format, and a particle setting of the wrong type that broke 9 of 16 scenes.
• In my new app, the frame on my desktop matched the frame Claude renders for itself exactly, a pixel difference of 0.
• My 'web wallpapers are slow' feeling had three causes: a 15 fps cap, the wrong graphics chip and slow effect code.

Sonnet 5.5 cost $249 to Opus 5.5's $255 over three long tests. Re-reading the conversation cost the same on both, and on one test Sonnet used twice the tokens.

• Effort mostly sets how much gets done. On the canal, Opus at Max did 9x Low's actions with 1.7x the thinking per action.
• Quality was close: all ten runs passed every test on the language task, and Opus won 4 of 5 head-to-heads on the canal.
• On the language task Max cost 16-18x Low for the same score. It clearly paid off only on the music videos.

After under an hour of training on my 8 GB laptop, the model drew Jefrie in 27 of 30 test prompts, where 998 renders of prompt tuning never got her once.

• Caption only what changes (pose, angle, setting) and her mask, grin, tail and cape soak into one trigger word.
• The first run did four images in 19 hours because a leftover settings file put it on the CPU. A guard now kills that.
• Misses: her day job (one training picture showed it), sleeping and a close crouch. It won't work on my other model.

The game ran on stand-in shapes 69 minutes after the first commit, but the two fighters took another full day, one job at a time through my laptop's one Blender.

• The final check found 75 problems, 3 of them blockers. One drew junk over the fight while all 40 automated tests passed.
• An AI 3D generator blurred BloodTailor's grin and fused his cloak to his arms, so an agent scripted him in Blender.
• 130 agents, all Opus 5.5, about $1,730 at API list prices. Most of the 4.5 billion tokens were context read again.

The video model copied my guide animation's movement as faithfully as its camera, so a badly animated guide made the motion worse than no guide at all.

• With the guide, the camera and the scene held together from shot to shot. The dome, towers and sandbags stayed put.
• Without it, the motion felt far better, but where Jefrie stood and which way her scythe swung were often unclear.
• Even a guide made of gray boxes and spheres dragged the motion down. Keep it simple enough that you can animate it well.

The image a video starts from sets its look more than the prompt. Three of nine AI video shots drifted to glossy 3D, and all three began on my smoother stills.

• Each step got four options, picked only once the next four existed. That lag turned Echo from witness to courier.
• Describing beats naming. Every model drew the Storm Bell landmark as a glass dome until I described an iron birdcage.
• Long laptop films drift from painted to smooth 3D by clip three. A painted frame every 2 to 4 clips pulls them back.

A script Claude wrote to run inside my game made a 2:45 music video with no footage or AI video. It worked because the game already had test tools a script could drive.

• Free-flying shots came out flat gray because free-fly mode turns off the fog. Dropping it showed nine cameras in rock.
• Gunshots, hits and moans are the game's own sounds, each on its frame. Nobody listened to the mix before it shipped.
• About 49 minutes on my laptop's graphics card. Most of the five hours went to waiting for a turn on it.

The picture got good because Claude could look at its own frames, while the music could only be measured and nobody listened to it before it shipped.

• All eight fix passes came from looking at frames. One found a flipped camera sign showing a 3D pagoda from underneath.
• The music took 81 s to render 30 s of sound until notes were built just before playing. Now it takes 4.5 s.
• With my character Jefrie, the AI video model copied its 3D guide's look. A gray guide plus painted end frames fixed it.

An agent that can only see its own session blames itself for any change on the machine, and here the retraction never reached the notes the confession did.

• The confessing session's 41 agents ran 843 commands and deleted nothing. The other one deleted the files, as I'd asked.
• The notes agent read only my first prompt to the deleting session, 'please do not delete anything', not the later one.
• One search of the deleting session's transcript for the file names would have settled it. Nobody ran it, including me.

Agent-built models walked within a day but weren't very good, and the generated one looked like her from every side but couldn't bend until an agent rebuilt it.

• GPT-6 Astra's model looked better than Claude's, but only from one angle. In a same-prompt walk-off I picked Claude's.
• From one picture, TRELLIS.2 made a fused 685,000-triangle shell in 321 s. Making it able to bend took 1 h 46 min.
• Her rules, like a tail heavier than she is, came from correcting twenty versions. My first prompt was one line.

Both builds passed their own tests and neither was playable, because I took out the one thing that made the first game work: me playing it between turns.

• Build one: about 220 agents, 588 passing tests, and a game that never unpaused. Every test went around the pause.
• Build two's autopilot only won after enemy grab damage went from 35 to 15 and healing doubled. No person ever played it.
• Pacing that held: about half the output that used to hit my usage limit, four agents at most. 15 hours, no limit stops.

Reading the transcripts of an 80-agent overnight run showed an agent that died four hours in after about $22 of usage, and I never noticed.

• It showed the review loop ran out of problems: 380 raised in round one, 126 in round two, 0 in round three.
• The trick was joining two files Claude Code writes: one says what each agent was for, the other what it did.
• It runs only on your computer and uploads nothing, and unknown numbers show as a dash with the reason, never as $0.

Before I hand work to AI, I ask if I need to be the one who did it. Tedium I already understand goes to the agent, and deciding what to build stays with me.

• One prompt cleared game work I'd burned out on for ten months. Then I handed over every goal and stopped recognizing it.
• I had AI build a viewer for my agents' logs before I knew what I needed, then spent hours of review working that out.
• AI amplifies the taste you have and doesn't install taste you don't, so I have to pick a few things to get deep in.

Finding useful things is cheap now and reading them isn't, so the layer worth building is a validator that keeps most findings from ever reaching you.

• Hundreds of agents, spun up from your meetings, conversations and screen, and almost none of them ever talk to you.
• Each would test an insight with a quick prototype first, and if it still isn't vital to you right now, file it unseen.
• It isn't far off: models get cheaper, smarter and better at coordinating, and now have real context to judge with.

After the rebuild, agents got 9 of 10 lookup questions right instead of 6, and the one confident wrong answer, caused by stacked-up corrections, was gone.

• Give every fact one home and fix mistakes in place. The old wrong answer came from a pile of added-on corrections.
• Pages of 750 to 4,000 words work better for agents than the usual advice of short, 250-word notes.
• Notion over files on my PC: no AI phone app can read a local folder, and all the big AI apps can connect to Notion.

Keeping agents busy was the wrong goal, because checking in every hour meant reloading the project in my head twenty times and never understanding it.

• The reviews added up to only 3 or 4 hours. The real cost was loading the project back into my head twenty times.
• Now I burst: one big run with lots of agents, then a review day of about four hours while nothing changes.
• Still unproven: I don't know the right size for a burst. The biggest so far ran 8 hours with 80 agents.

When I aimed AI at learning it taught me to program, and when I aimed it at building I built more and lost track of my code.

• What I learn now is operating: which tool, which model, how hard to let it think, and what can run without me.
• Give agents fewer rules than you think, but state what you both take for granted, or every agent works it out again.
• Still open: if I never build another foundation by hand, will I notice when the model is wrecking one?

I think the week worked because the model built straight from specs I'd written by hand before any code, documents I'd expected to be worthless.

• The fix for a flickering screen was moving its live data out of React, the library that draws it, to a separate store.
• 137 commits in four days built every requirement I'd ever written. A month later I'm still reviewing it all.
• Large, tedious, risky and nothing new: any one is survivable, but all four together will end a solo project.

Never hand the model a decision you can't check, because it won't warn you about a bad name, a bloated design, its memory filling up or a file it never saw.

• As a long chat fills its memory, it gets dumber, more agreeable and more likely to make things up, with no warning.
• It agreed with every guess, even after I told it I was a novice. A question would have cost ten seconds.
• Each new component copied every option of the one before plus one more, until my form had 25 of them.

The hard part of an AI app is everything around the model call, like queueing requests, chaining steps, handling failures and telling the user what's going on.

• The model kept breaking my code's structure, so I read every line and learned that structure myself.
• Picking by hand which files Claude saw, on every message, forced me to learn my own code.
• One GPU runs one prompt at a time, so every request waits in a queue and the whole app is built to wait.

The model wrote all my code, but it couldn't set up my computer or tell me what I didn't know to ask, and that was the real beginner barrier.

• I didn't know where files go, had no code editor and had never used a terminal, so GitHub was a wall.
• I did git, the tool that tracks code versions, by hand and badly. Much later I learned a code editor does it for me.
• Each gap was invisible until I ran into it. Then one sentence from the model fixed it for good.

From one vague message, Claude pulled my original skin off Mojang's servers and worked out my own old account had been holding my name the whole time.

• The old account is almost certainly tied to an old email I can't get into.
• My old worlds were never stored online. They'd only survive as a backup in my own .minecraft saves folder.
• From 'an upside-down snowcone' it named Mineplex's Blood Helix effect and rebuilt it as a datapack you can download.

Left to pick, the prep session ran everything on the strongest model and used 30% of my weekly limit for it in one evening, so I'd choose the models myself.

• The overnight build took 8h39m, with one lead agent on Fable 5, Claude's top model, steering 80 agents, 76 on Opus 5.
• The blind critics flagged 380 problems in round one, 126 in round two and 0 in round three.
• Three sessions with no shared memory ran as a relay, each writing the next one's prompt, from one prompt of mine.

Having the agent take screenshots while it played and narrate them at the end gave me full context on a two-hour session in eleven minutes.

• It found five real bugs by playing and fixed all five. The worst silently ate any message over 500 characters.
• The screenshots had to be taken during the work. Asked for afterwards, none of them would exist.
• Its review: the story is the strongest system, and the AI referee is too kind. It never once said no to a good speech.

Asking for six districts instead of two made the whole game worse, even the hero's lines, the shop jokes and the camera, which I hadn't changed.

• One prompt holds about fifteen minutes of game, two districts, a few systems and one joke per location.
• To check if you asked for too much, look at something you didn't change and see if it still works.
• I gave it the shape of every joke and it wrote funnier lines than I would have.

Claude couldn't fix its own 3D bugs even with screenshots, but it built me an editor to fix them myself, so on 3D projects I now ask for tools and tests first.

• It worked hard on code but rationed effort on the story, so I had to push every beat of the lore by hand.
• AI voice lines and music from ElevenLabs made the rough graphics feel alive. On a small game, do sound first.
• Lyrics that must match the game's lore should come from the lore document. For a feeling, let the music AI write them.

What I'd figured out about AI lived in scratch files and my head, useless to me in six months and invisible to everyone else. So now it goes here.

• Entries are short and unpolished, filed as I go.
• Some will be how I fixed something. Some will be half-formed ideas I'm still testing.