Kagi retweeted
THIS IS F*CKING GOLD
Anthropic just shared a Claude Code trick: 3 models, 1 command. Opus judgment on the big calls, cheaper models on everything else
update first: claude update
then launch it like this:
claude --model sonnet --advisor opus
the split:
→ Sonnet 5.5 runs the session and writes the code
→ Haiku 5.5 reads files and docs in parallel
→ Opus 5.5 waits on call as the advisor
→ Opus stays quiet on routine turns
you tell Claude to call Opus at 3 moments:
→ before the plan locks
→ when the same error shows up twice
→ before the task is called done
Sonnet builds, Haiku explores, Opus reviews
the full tree:
> Sonnet 5.5: main session, edits code
> search: maps the repo, Haiku 5.5
> grep: finds call sites, Haiku 5.5
> docs: pulls the API reference, Haiku 5.5
> Opus 5.5: advisor, reads the whole transcript when asked
paste this into Claude Code:
"Rebuild my setup around this tree:
1. main session on Sonnet 5.5, Opus 5.5 as advisor
2. three read-only Haiku 5.5 subagents: search (maps the repo), grep (finds call sites), docs (pulls the API reference)
3. one rule in CLAUDE.md: consult the advisor before locking a plan, after the same error twice, and before calling a task done
Show every change as a diff first. No edits until I say go."
the advisor is experimental. it runs on the Anthropic API only, not on Bedrock, Google Cloud or Foundry
if you run Opus on every turn, you pay Opus rates for file reads too. save this before your next big refactor
Kagi retweeted
BREAKING: Anthropic's docs let you put Opus 5.5 in charge of 3 Sonnet 5.5 builders
a dev on X shared the setup. 45K views in 2 days
most people still ask 1 agent to do everything. it plans, codes and tests in one chat, and the chat gets messy
here is the team:
→ 1 lead on Opus 5.5 plans the work and hands out tasks
→ 3 builders on Sonnet 5.5: frontend, backend, tests
→ each builder works in its own context window
→ 1 shared task list tracks who owns what
→ builders message each other when their parts must match
→ the lead steps in at 3 points: before the plan locks, when a builder is stuck, before done
start with 1 line:
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 claude --model opus
the catch: it is experimental. Anthropic says a team uses a lot more tokens than 1 session. and 2 builders on the same file overwrite each other, so give each builder its own files
1 messy chat → 1 lead + 3 builders → 1 shared task list
the prompt and the docs link are in the first comment ↓
your 1 agent is still doing a whole team's job alone
Kagi retweeted
THIS AI PODCAST MADE $15,248 AND NEITHER HOST IS A REAL PERSON 🤯
no camera. no mic. no studio. no guests to chase down.
here's roughly how the setup works:
1. ask an AI chatbot for a viral podcast idea + the opening lines
2. paste the script into an AI video tool that has a "podcast" template
3. it generates two hosts who look real and talk to each other
4. post on YouTube, repeat, let the views turn into revenue
the crazy part is i honestly couldn't tell at first. the lighting, the mic, the way they react to each other... it all looks like a normal studio podcast
and that's what bugs me a bit tbh. if a faceless channel can pull $15k with people who don't exist, what does "authentic content" even mean in 2026?
but also, the barrier to start a show is basically zero now. one person + a laptop = a whole podcast studio
would you watch an AI podcast if you knew the hosts weren't real? or would you click away the second you found out? 👇
BREAKING: Anthropic's docs let you put Opus 5.5 in charge of 3 Sonnet 5.5 builders
a dev on X shared the setup. 45K views in 2 days
most people still ask 1 agent to do everything. it plans, codes and tests in one chat, and the chat gets messy
here is the team:
→ 1 lead on Opus 5.5 plans the work and hands out tasks
→ 3 builders on Sonnet 5.5: frontend, backend, tests
→ each builder works in its own context window
→ 1 shared task list tracks who owns what
→ builders message each other when their parts must match
→ the lead steps in at 3 points: before the plan locks, when a builder is stuck, before done
start with 1 line:
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 claude --model opus
the catch: it is experimental. Anthropic says a team uses a lot more tokens than 1 session. and 2 builders on the same file overwrite each other, so give each builder its own files
1 messy chat → 1 lead + 3 builders → 1 shared task list
the prompt and the docs link are in the first comment ↓
your 1 agent is still doing a whole team's job alone
docs: code.claude.com/docs/en/agen…
Run this session as team lead on Opus 5.5.
Spawn 3 teammates on Sonnet 5.5: frontend, backend, tests.
Give each teammate its own files. Shared schema and config stay with you.
Use one shared task list for owners, progress and blockers.
Let teammates message each other when their parts must match.
Step in before the plan locks, when a teammate is stuck, and before done.
Show me the plan first. No edits until I say go.
Kagi retweeted
BREAKING: WALL STREET spends millions on quant teams.
I spent $20 on Anthropic.
Last night I gave it a paper account.
$10,000. Fake money. One prompt.
I typed:
"Analyze profitable Polymarket wallets from the last 90 days and build me an edge."
Then went to sleep.
Woke up to:
+$4,237.
(simulated run)
While I slept, the agent:
• scanned 10,000 wallets
• filtered fake alpha from real edge
• tracked sizing, timing, and conviction
• found traders consistently beating the market
Then it built a fully automated strategy:
→ monitors breaking news
→ maps narratives to active markets
→ detects mispriced odds
→ executes arbitrage
→ sizes positions dynamically
Started at 11:47 PM.
By morning:
94 trades.
Zero manual input.
No fear.
No greed.
No second-guessing.
Just data reacting faster than humans can.
People keep asking if AI will replace jobs.
Wrong question.
The real question is:
what happens when one person gets tools that used to belong only to Wall Street?
You only need Anthropic + laptop + 1 hour/day.
Paper first.
Most people skip it and test with their rent.
Kagi retweeted
BREAKING: an OpenAI co-founder's notes setup is at 22M views and 108K bookmarks. this is f*cking wild
Andrej Karpathy now spends a big part of his tokens on notes, not code
most people still google strangers. he asks his own wiki
the setup takes about 5 minutes:
→ brew install --cask obsidian, then mkdir ~/vault
→ open Claude Code in ~/vault and ask for a CLAUDE.md: raw/ for sources, wiki/ for pages, plus ingest and lint
→ drop a pdf, a transcript or meeting notes into raw/ and type: ingest
→ the model writes the wiki pages and backlinks. Karpathy says he rarely edits them himself
→ ask questions inside the vault. answers point to your own pages
→ save good answers back into the wiki, so every question adds up
→ type: lint. it finds pages that disagree and gaps to fill
honest part: he says it works at a ~small scale. his wiki is ~100 articles and ~400K words, and he still vibe coded a small search engine on top
108K people bookmarked his post. this time put it in raw/
full guide in the article below
BREAKING: an OpenAI co-founder's notes setup is at 22M views and 108K bookmarks. this is f*cking wild
Andrej Karpathy now spends a big part of his tokens on notes, not code
most people still google strangers. he asks his own wiki
the setup takes about 5 minutes:
→ brew install --cask obsidian, then mkdir ~/vault
→ open Claude Code in ~/vault and ask for a CLAUDE.md: raw/ for sources, wiki/ for pages, plus ingest and lint
→ drop a pdf, a transcript or meeting notes into raw/ and type: ingest
→ the model writes the wiki pages and backlinks. Karpathy says he rarely edits them himself
→ ask questions inside the vault. answers point to your own pages
→ save good answers back into the wiki, so every question adds up
→ type: lint. it finds pages that disagree and gaps to fill
honest part: he says it works at a ~small scale. his wiki is ~100 articles and ~400K words, and he still vibe coded a small search engine on top
108K people bookmarked his post. this time put it in raw/
full guide in the article below
Kagi retweeted
i still don't get why people build a second brain for Claude Code from scratch
7 repos already did it. 75,000+ stars combined
1. obsidian-skills: 6 skills from the CEO of Obsidian. teaches Claude its formats and tools
github.com/kepano/obsidian-s…
2. claude-obsidian: drop text in inbox/, check the plan, get linked notes
github.com/AgriciDaniel/clau…
3. obsidian-second-brain: saves decisions and tasks, loads them back next session
github.com/eugeniughelbur/ob…
4. karpathy-llm-wiki: unofficial Karpathy-style wiki. ingest, query, lint
github.com/Astro-Han/karpath…
5. second-brain-os: starter vault with 18 skills and 72 commands
github.com/undefined-ui/seco…
6. ai-second-brain: turns your ChatGPT and Claude exports into linked notes. run only step 1 of the README
github.com/charlie947/ai-sec…
7. llm-wiki: rules load every session, wiki pages only when needed. skip its tip to keep secrets in memory
github.com/MehmetGoekce/llm-…
start with #1. it takes 2 commands:
/plugin marketplace add kepano/obsidian-skills
/plugin install obsidian@obsidian-skills
all 7 run on plain markdown. you can open every file in any editor
full guide in the article below
Kagi retweeted
holy sh*t
I gave Anthropic one task:
"Look at every top Polymarket wallet from the last 90 days and build me something"
Then I went to sleep.
Not a script.
Not an if-then bot.
An agent that reads live news,
matches it to markets,
spots wrong prices,
and sizes every bet with Kelly.
I started it tonight.
Closed the laptop.
No input.
No second-guessing.
Wall Street pays millions for this.
I pay $20 a month.
Kagi retweeted
A $5 CABLE + CLAUDE CODE ( FREE VERSION ). YOUR CAR JUST GOT HACKED. NO CAP.
This guy plugged a cheap OBD2 adapter into his dad's car and gave Claude Code one job:
"Figure out how to talk to this car and build me a live dashboard."
5 minutes later, things got pretty damn interesting.
Real engine data. Live temperatures. Diagnostic readings. All flowing into an app Claude helped build from scratch.
Here's what was actually happening under the hood.
Claude wasn't just drawing some fancy dashboard. It had to figure out how to communicate with a real car.
And the tech behind it is surprisingly simple:
→ ELM327 + OBD-II: A cheap adapter connects the laptop to the car's diagnostic interface, giving access to supported ECU data.
→ SAE J1979: Commands like 010C request engine RPM, 0105 reads coolant temperature, and 010D checks vehicle speed.
→ Python + python-OBD: Sends diagnostic queries, reads sensor values and converts raw responses into actual measurements.
→ pySerial: Handles communication with compatible adapters, including connection errors and timeouts.
→ Claude Code: Helps write the connection logic, fix broken scripts and turn incoming vehicle data into a working application.
→ HTML + CSS + JavaScript: Can transform those numbers into live gauges, temperature indicators and a custom dashboard.
→ DTC diagnostics: Reads supported fault codes so you can understand what that annoying check-engine light is complaining about.
And here's the funny part.
You don't need 500 lines of code just to ask your engine what it's doing.
A basic read-only query looks like this:
```python
import obd
car = obd.OBD()
rpm = car.query(obd.commands.RPM)
temp = car.query(obd.commands.COOLANT_TEMP)
print(rpm.value, temp.value)
```
A few lines of Python. Real engine data.
The dashboard shows 69°C coolant temperature, battery voltage and diagnostic status.
And the cool part? Claude can customize everything.
Racing-style RPM gauges? Done.
Engine fault explanations? Easy.
A full Knight Rider-style dashboard? Hell yeah, lol.
But here's the twist.
He didn't bypass the car's security. He made it communicate with custom AI-built software.
5 minutes. One prompt. A $5 cable.
Next time your check-engine light comes on, imagine opening your own diagnostic app instead of guessing what's wrong.
Authored by @Vectorxeth
Kagi retweeted
THIS 17 Y/O MADE $2,319 IN ONE TRADE WITH AI. HERE'S WHAT THE VIDEO DOESN'T SHOW YOU
the "AI analyzer" he uses is not magic. it's a vision model reading a screenshot. you can do the exact same thing for free in 2 minutes
how to do it yourself:
1-open TradingView (free), pick any chart
2-screenshot it
3-drop it into ChatGPT, Claude or Gemini and ask: "act as a technical analyst. identify trend, support/resistance, and give me an entry, stop loss and take profit with your reasoning"
4-that's it. that's the whole "AI analyzer" you saw
now the part nobody puts in these videos:
AI reads patterns, it doesn't see the future. same screenshot can give you a different answer twice
you see the one +$2,319 trade. you don't see the 10 that hit the stop loss
most retail day traders lose money. AI doesn't change that math, it just makes losing faster
if you still want to try it: paper trade for a month first. TradingView has it built in. if AI setups don't win on fake money, they won't win on real money
so where's the real money here?
the people selling the "AI analyzer" subscription. a vision model + a nice UI + $49/month
the people making these videos. a $2k screenshot gets millions of views, views turn into sponsors and affiliate deals
building tools: a simple "upload chart - get analysis" app is a weekend project now
the trader takes the risk. the tool builder gets paid every month either way
which side are you on?
i still don't get why people build a second brain for Claude Code from scratch
7 repos already did it. 75,000+ stars combined
1. obsidian-skills: 6 skills from the CEO of Obsidian. teaches Claude its formats and tools
github.com/kepano/obsidian-s…
2. claude-obsidian: drop text in inbox/, check the plan, get linked notes
github.com/AgriciDaniel/clau…
3. obsidian-second-brain: saves decisions and tasks, loads them back next session
github.com/eugeniughelbur/ob…
4. karpathy-llm-wiki: unofficial Karpathy-style wiki. ingest, query, lint
github.com/Astro-Han/karpath…
5. second-brain-os: starter vault with 18 skills and 72 commands
github.com/undefined-ui/seco…
6. ai-second-brain: turns your ChatGPT and Claude exports into linked notes. run only step 1 of the README
github.com/charlie947/ai-sec…
7. llm-wiki: rules load every session, wiki pages only when needed. skip its tip to keep secrets in memory
github.com/MehmetGoekce/llm-…
start with #1. it takes 2 commands:
/plugin marketplace add kepano/obsidian-skills
/plugin install obsidian@obsidian-skills
all 7 run on plain markdown. you can open every file in any editor
full guide in the article below
Kagi retweeted
it's getting f*cking easy to give AI a memory
an OpenAI co-founder let it write him a wiki: 100 articles, 400K words
he almost never edits it. the AI writes every page
mkdir -p vault/raw vault/wiki && cd vault && claude
the setup, step by step:
→ drop articles, papers, repos and images into raw/
→ type: "compile raw/ into a wiki. one page per idea, with backlinks"
→ it writes the pages and keeps its own index, so at ~100 articles he needed no fancy search
→ ask your questions here, it answers from the wiki
→ good answers get filed back as new pages, so every question makes the wiki bigger
→ run a health check: it finds facts that don't match and fills the gaps
→ open the folder in Obsidian to see every page and link
setup: 1 folder, 2 subfolders, 1 command
this is his setup at ~100 articles. he calls his own tooling a hacky collection of scripts
most people open a new chat every time. he keeps adding to one wiki
the video is a mock of a vault. the 100 articles and 400K words in it are his numbers, not mine
his questions add up. yours are gone when you close the tab
i show you how to let AI do your work, daily.
it's getting f*cking easy to give AI a memory
an OpenAI co-founder let it write him a wiki: 100 articles, 400K words
he almost never edits it. the AI writes every page
mkdir -p vault/raw vault/wiki && cd vault && claude
the setup, step by step:
→ drop articles, papers, repos and images into raw/
→ type: "compile raw/ into a wiki. one page per idea, with backlinks"
→ it writes the pages and keeps its own index, so at ~100 articles he needed no fancy search
→ ask your questions here, it answers from the wiki
→ good answers get filed back as new pages, so every question makes the wiki bigger
→ run a health check: it finds facts that don't match and fills the gaps
→ open the folder in Obsidian to see every page and link
setup: 1 folder, 2 subfolders, 1 command
this is his setup at ~100 articles. he calls his own tooling a hacky collection of scripts
most people open a new chat every time. he keeps adding to one wiki
the video is a mock of a vault. the 100 articles and 400K words in it are his numbers, not mine
his questions add up. yours are gone when you close the tab
i show you how to let AI do your work, daily.
Kagi retweeted
What the f*ck did Claude Opus 5.5, Gemini 4 Argon and GPT-6.1 Terra just make
I asked them to recreate the legendary cow video.
Opus gave me a jelly cow sprinting on its udder through a sunny farm.
Gemini gave me the same cow, but shinier.
GPT gave me a f*cking coffee table on pink stilts.
Only one of them didn't make it over the fence. Which one?
Kagi retweeted
this is f*cking wild.
i simulated a polymarket copy bot for 27 days. $200 in paper money became $11,460.
the whole stack:
• Anthropic's coding agent
• 4 open-source repos
• $25 a month
step 1 - find the smart money
one prompt: "find wallets with 100+ trades and 70%+ win rate. rank by profit."
14,212 wallets scanned. 47 kept.
step 2 - cut the markets
a Rust CLI pulls 500 live markets. it checks spread, liquidity, timing and whale moves.
500 became 35. 93% rejected before i looked.
step 3 - 3 agents vote
arbitrage, convergence, whale copy. no shared memory.
2 agree = full size. 1 agrees = half size. 0 = no trade.
this cut 40% of the losing trades.
step 4 - leave early
91% of smart wallets exit before the end.
so the bot exits first, at 85% of the expected move.
after 27 days:
271 trades
74.2% win rate
sharpe 2.47
max drawdown -$61
it still took a -$23 loss on Oil $90. the agents said full size. remember that part.
most people show you the green line. nobody shows the 214 trades the bot refused to take.
the stack costs less than Netflix.
next run: 1,000 markets instead of 500. what breaks first?
Kagi retweeted
it's getting f*cking cheap to run Claude Code. one flag puts Sonnet 5.5 on the code and calls Opus 5.5 in for 3 decisions
claude --model sonnet --advisor opus
Opus stays on call and writes 0 lines of code
→ launch the session with that flag, Sonnet 5.5 takes every routine turn
→ hand the searching to 3 Haiku 5.5 subagents: search, grep, docs
→ Sonnet usually calls Opus before a plan locks
→ it calls Opus again when the same error comes back twice
→ and a 3rd time before it says "done"
→ Opus reads the full session each time, every tool call and every result, and sends advice back
→ add 1 line to your prompt for more checks: "consult the advisor before you lock a plan"
every Opus call re-reads the whole session from zero, with no cache. a long session with many calls eats the savings, remember that part
it's also experimental and runs only on Anthropic's API, not on Bedrock or Google Cloud
most setups run Opus on every turn, file reads included
$0 extra tools. on a subscription, advisor tokens count toward the plan you already pay for
the video shows a mock session with 41 turns. 3 of them go to Opus
how many of your turns today needed Opus?
i show you how to let AI do your work, daily.
Kagi retweeted
9 prompts. 1 skill. a six-legged crane walks, lifts, grasps and comes apart into labeled parts, and every move reads one 2,400 rpm input. you are still keyframing one gear by hand
the skill is Elvar's teardown. free, runs in Claude Code, ships the whole machine as one offline HTML page
the rule inside it is simple. every part is a function of one input. on this board that input is a V8 power pack at 2,400 rpm. the crankshaft with 8 pistons, the 96-tooth slewing ring, 6 leg drives, the boom and the hook all read the same model. so walking, lifting, grasping, the exploded view and the lidar sweep are one simulator in six modes, not six animations glued together
here is the gist:
→ type /teardown, name the machine, attach one real reference. without it you get a toy
→ split it into assemblies you can isolate: hull, boom, legs, manipulator, optics
→ give every part a real number before it gets a shape: teeth, sections, drives, pistons
→ drive everything from that one input
→ add modes as views of the same model: walk, lift, grasp, explode, scan
→ press any part and it slides out with its label while the rest steps aside
→ play it at 1/20 speed and watch for the one leg that is out of step
six modes, one model. fix one number and every mode fixes with it
the 9 prompts are in the quoted article
legs on a tripod gait or a boom slewing 360° on 96 teeth, which assembly breaks first when you build yours?
↳ concept board, crane is illustrative, method from the article
