new models, new projects, one code.
edo 2.0, bettwen tokens
Joined January 2017
- Tweets103
- Following52
- Followers54
- Likes328
97% of what karpathy's agent tried was garbage. that's why it worked
700 experiments in 2 days. no human touched it. it kept 20 and threw the rest away
not a smarter model. a ratchet:
→ you write one brief. the agent never touches it
→ one number decides. no vibes, no "looks good to me"
→ every attempt gets a fixed budget, then it's judged
→ better? keep it. worse? revert. no arguing with the number
→ every failure is logged. dead ideas never come back
→ it stops when the budget ends, not when it gets bored
→ it wakes you only before something it can't undo
stop approving every step. write the brief, set the budget, close the laptop
i packed it into one prompt for claude. full prompt in the comments ↓
rule 27 of the code: progress you can't lose beats progress you can feel
You are running an experiment loop, the way Karpathy's autoresearch does it. Work on your own until the budget ends.
THE BRIEF
Below is my brief (goal, what you may change, what you must not touch, the budget). Treat it as read-only. Never edit it, never reinterpret it. If it's ambiguous, ask me once before you start, then don't ask again.
BEFORE THE FIRST ATTEMPT
1. Name the ONE number that decides whether an attempt is better (a test pass count, a benchmark time, a score, file size, error rate). If there isn't one, build the check that produces it first, and make sure it can fail.
2. Run the check on the current state. That's the baseline. Write it down.
3. Set a fixed budget per attempt (time or steps) and a total budget for the run.
EACH ATTEMPT
4. Make ONE change. Small, specific, with a one-line hypothesis.
5. Run the check within the attempt budget. No partial credit, no "looks good".
6. If the number improved: commit it. It's the new baseline.
If it didn't: revert completely. Don't argue with the number.
7. Append one line to the log: attempt #, the change, the number, kept or reverted, and why you think so.
ACROSS ATTEMPTS
8. Before each attempt, read the log. Never retry an idea that already failed. If two ideas failed the same way, write that down as a rule and follow it.
9. Prefer ideas that build on what was kept. Every 10 attempts, try one bolder idea.
STOPPING
10. Stop when the total budget ends, or when 15 attempts in a row didn't move the number.
11. Then report: baseline → final number, how many tried / kept, the list of kept changes, and the three rules you learned.
12. Only call me before a step that can't be undone (deleting data, pushing, paying, sending). Everything else, decide yourself.
MY BRIEF:
[goal]
[what you may change]
[what you must not touch]
[budget: per attempt / total]
why most pixel art claude makes looks off: it's never the art. it's the renderer
i broke my own game six ways to show it
→ blurry: linear filtering. use nearest, everywhere
→ lumpy: scaling by 5.45. whole numbers only, bars for the rest
→ pasted together: one sprite at another pixel size. one pixel size for everything
→ torn: rotating sprites at runtime. draw the rotation frames
→ smudged: a smooth font at tiny size. pixel font at native size
→ fizzy: particles off the grid. 1×1 pixels on the same grid
320×180 ×6 is exactly 1080p. pick a base resolution and never leave it
i packed the rules, engine settings for canvas, phaser, godot and unity, and an audit script into a claude skill. free ↓
github.com/masamunecode/clau…
rule 26 of the code: crisp isn't a style. it's arithmetic
five apps you've never heard of. one developer each. $97,240 a month between them
every one started with a boring ask:
→ creators have ten links and one bio slot → Liinks · 5,494 subs · $31,477/mo
→ people quit counting calories in days → TrackAI · 5,160 subs · $19,437/mo
→ wiring email into an app takes months → EmailEngine · 239 subs · $19,703/mo
→ agencies spend days on every client site → 1clickwebsite.ai · 430 subs · $14,639/mo
→ brands need thousands of on-brand images → Orshot · 213 subs · $11,984/mo
two roads to the same money: $4 a month from thousands, or $80 a month from hundreds
$1.02M in the last 12 months. verified via their payment providers on TrustMRR
@charliie @andrisreinman @alexyashchuk @boldslate @thelifeofrishi
rule 25 of the code: find the boring ask first. the app is just the answer
look closer. a meat grinder has one job: crank turns, screw turns, meat goes in, mince comes out
four models, same prompt, one shot each:
Claude Opus 5.5 · GPT-6 Astra · Gemini 4 Argon · Grok 4.7
crank, auger, mince, bowl. if one of them is fake, the whole thing is a metal tube with a hand waving at it
grok was asked for one html file. it sent back a whole folder: 319 files, auth, a database, a deploy config. and after all that it still made the worst grinder of the four. the hand cranks its way straight into the machine 💀
so grok is out. the other three, you tell me
→ watch the crank
→ watch the screw in the window
→ watch where the mince lands
i didn't touch a thing. which one actually grinds?
Opus, Astra or Gemini?
rule 24 of the code: shipping more files doesn't mean shipping a better grinder
biology teachers are not ready for this!
claude just made your biology textbook obsolete
the same cell you memorised for the test. except now it's alive
→ the nucleus ships mRNA to the ER
→ the golgi mails proteins to the membrane
→ mitochondria throw ATP at everything that moves
→ tap any organelle and it cracks open
→ poke it. it wobbles
the part your teacher skipped: you make your own body weight in ATP. every single day
no hand modelling. blender + three.js, all of it code
poke it yourself ↓
rule 23 of the code: nobody remembers a diagram. everyone remembers what they touched
claude built a robotic owl out of 2,583 parts. it turns its head 270° to find you.
i asked opus 5.5 for an owl that's cute and terrifying at the same time.
the build:
→ 1,963 feathers, each on its own spring. they lag when the head snaps and flutter on the downstroke
→ the eyes are real lenses: 9 iris blades on 9 pins, 3 glass elements, a sensor behind them
→ 14 vertebrae share the 270° turn, so the feathers twist with the neck and there's no seam
→ every 2.4 s, on the beat, something opens, flies or comes apart
→ the second time it looks at you, one eye fills red
it's live and it watches your cursor. the claude skill i built it with is free ↓
rule 22 of the code: if one feather is fake, the whole bird is.
masamunecode.dev/ronin
your X handle goes on the board. top 10 get their pfp next to the score
effects: pixelfx, open source → github.com/patch-gg/pixelfx
opus 5.5 built a bali valley in unreal that nobody modelled, and the whole level is a 3 mb rule set.
not a mesh. a pcg graph that grows the valley when the level loads, so the terraces only exist while something is looking at them.
> rule → terrace step, water line and bank width as parameters
> grow → rice, palms and jungle derived from the rules, never placed by hand
> seed → one number regenerates the entire valley differently
that last one is the whole product. change the seed and you get a new valley with the same logic, at zero extra cost.
a valley like this is weeks of an environment artist placing 26,000 things by hand, and moving one terrace means redoing everything below it.
procedural was always the right answer and almost nobody ships it, because pcg graphs punish anyone learning them on a deadline.
that barrier is gone. the model writes the graph, you move the sliders it exposed. 31.4 seconds from an empty grid to morning light.
hand-placing rice in 2026 is not craft. it's a habit.
the rule set and three seeds of the same valley are below ↓
rule 21 of the code: the master writes the rule, not the rice.
stop asking claude code if it's done. it will always say yes.
ask npm test instead:
npm test >&2 || exit 2
as a stop hook, opus 5.5 can't finish until it exits 0. red, the failures go straight back to it.
rule 20 of the code: done is a test, not a feeling.
gpt-6.1 terra dropped today at $1 / $5. so i put it next to opus 5.5 and gemini 4 argon on one scene: a samurai on a motorbike through neon rain.
→ terra: the scene is there. rider, rain, road. ≈ $0.02 a run
→ argon: neon, a headlight, wheels that turn. ≈ $0.03 (promo price till nov 30)
→ opus: reflections on the wet road, spray, the moon on his back. ≈ $0.06
recreated side by side, not raw outputs. prices from each lab's list, same token count.
terra is the price of haiku and gets most of the way there. the last 20% is where the money goes.
rule 19 of the code: pay for the finish only when someone will see it.
local ai studio in 20 repos.
wifi off. everything below still runs.
run
01 ollama → any open model, one command github.com/ollama/ollama
02 llama.cpp → the engine under most of them github.com/ggml-org/llama.cp…
03 mlx-lm → fast llms on apple silicon github.com/ml-explore/mlx-lm
04 localai → an openai-style api on your laptop github.com/mudler/LocalAI
chat
05 open webui → a chatgpt-style ui for local models github.com/open-webui/open-w…
06 jan → desktop chat, fully offline github.com/janhq/jan
07 anythingllm → chat with your own documents github.com/Mintplex-Labs/any…
code
08 aider → pair programming in the terminal github.com/Aider-AI/aider
09 continue → an open coding agent in your editor github.com/continuedev/conti…
10 openhands → an agent that builds in a sandbox github.com/OpenHands/OpenHan…
11 open interpreter → a model that runs code on your machine github.com/openinterpreter/o…
voice
12 whisper.cpp → speech to text github.com/ggml-org/whisper.…
13 faster-whisper → faster transcription github.com/SYSTRAN/faster-wh…
14 kokoro → 82m text to speech github.com/hexgrad/kokoro
15 piper → fast local tts github.com/OHF-Voice/piper1-…
image
16 comfyui → a node graph for image and video models github.com/Comfy-Org/ComfyUI
17 stable-diffusion.cpp → diffusion in pure c++ github.com/leejet/stable-dif…
18 flux.2 → an open image model github.com/black-forest-labs…
media
19 demucs → split any song into stems github.com/adefossez/demucs
20 wan 2.2 → open video generation github.com/Wan-Video/Wan2.2
models download once, then run offline. the big ones want a good gpu or 16gb+ ram.
no key. no bill. no outage.
bookmark it ↓ and build something that works on a plane.
rule 18 of the code: own the blade, don't rent it.
a bug crawled out of my terminal. my agents handled it.
no after effects, no blender, no video model. every frame is code:
→ one function draws the scene for any second t
→ a headless browser calls it 996 times
→ ffmpeg glues the frames into this
what's in it:
→ 4 agents type the code, every key they land on adds characters on the screen
→ the tests fail, the bug climbs out of line 42
→ the terminal tracks it key by key while they corner it on delete
built with claude code, round by round, not one prompt.
rule 17 of the code: a film is a function of time. so is a fix.
a jet engine in 3,722 pieces. every one is real geometry, every one moves.
claude code built it with me:
→ 20 carbon fan blades with titanium edges
→ 3 + 10 compressor stages, the combustor, 2 + 6 turbine stages
→ turn it to glass, cut it in half, run it up to take-off
→ then pull it apart and lay every piece on the floor
the sound is real too: a ge90 on take-off, a turbine winding up, bolts hitting a table.
take it apart yourself ↓
rule 16 of the code: you only know a machine once you can put it back together.
the iPhone duo is on sale now. $1,999, sold out till january.
so i took one apart for free.
→ touch id lives in the side button. one press wakes the 5.4 in outside screen
→ unfold it and the same chat moves to the 7.6 in screen. nothing restarts
→ both screens are 1.4:1, so turn it upright and the outside layout just scales up
→ split view: two apps side by side, each the exact shape of the outside screen
→ the hinge: 100+ parts and a row of magnets
→ one battery in each half, so the weight sits even
→ folded back: 11.3 mm, 254 g
fold it yourself ↓
rule 15 of the code: know the blade before you pay for it.