Drerki

Analyzing viral AI setups. What delivers results, what's just noise, and the real price tag.

[Drerki @Drerki]

ANTHROPIC JUST DROPPED THE CHEAPEST WAY TO RUN OPUS-LEVEL LOGIC

let Haiku drive the session, keep Opus on call for decisions that actually matter

the setup:
claude --model haiku --advisor opus

here is how the split works →

haiku handles:
→ writes code and runs tests
→ searches the repo in parallel subagents
→ pulls docs and finds call sites
→ drives the whole session at low cost

opus steps in at three moments:
→ right before a plan locks
→ when the same error hits twice
→ right before marking done

haiku builds. haiku explores. opus decides.

one level down: JEV routes micro-forks in milliseconds. tool choice, retry or stop, which file to open. the fast models never see these.

costs more than haiku alone, less than running sonnet or opus for the whole session.

needs claude code v2.1.293 or later, run claude update first

paste this into claude code:

"set up around this tree: run main session on haiku 5.5 with opus 5.5 as advisor. create three read-only haiku subagents: search maps the repo, grep finds call sites, docs pulls API reference. add one rule to CLAUDE.md: consult advisor before plan locks, after same error twice, before task marked done. route fast forks through JEV. show every change as diff first."

this is the cheapest opus you will ever run

[Addy Osmani @addyosmani]

https://t.co/2gyp2Erpxx

[Drerki @Drerki]

GTA 6 IS NOT OUT. THIS ONE WAS BUILT IN 134 MINUTES

one prompt. one $200 plan. no game engine, no assets, no audio files

what came back:
→ a character who sprints, jumps, kicks, and reloads a pistol
→ a wanted system that stacks to 5 stars, and cops who shoot back
→ any car is stealable. 3 gears, handbrake, live traffic
→ a garage with 10 paint jobs, 5 wheel sets, rim colors and neon underglow
→ a 9 checkpoint race on a timer, plus a courier run that pays $600
→ 9 districts, day into neon night, a minimap, hidden cash, a pelican on the pier

every gunshot, engine and siren is generated in python. zero audio files

the prompt was not clever. it was a job description:

"you're a senior game dev. build a third person open world crime game in a sunny 80s miami style city. make the loop tight: walk, fight, steal a car, get chased, earn money, spend it. every system feeds the next. cars need weight: gears, skids, real crashes. no audio files, generate every sound in code. after each system, play it yourself, screenshot it and fix what looks wrong. don't stop until a stranger could play for 10 minutes without hitting a bug"

that last line is the trick. it made the model test its own work

an indie studio quotes 8 months and $90,000. this took 2 hours and 14 minutes

the part nobody says: it is one district loop, not a game. and it still needs a human to check it

[Mika @mika_systems]

Claude Code + Opus 5.5 is godmode for building your own Claude Duet

it reads your inbox and calendar before you wake up. drafts every reply. sends nothing without your yes

it keeps working while you sleep

full course, step by step: https://t.co/bWBuLd8Gpg

[Drerki @Drerki]

@aksenHQ it lists folders, not files. one line per folder, not per note

when a folder gets crowded you add a subfolder line. the map grows by folders, so it stays tiny

the rule inside the map is "look in notes for facts", not a list of every fact

[Drerki @Drerki]

CLAUDE ANSWERED "I DON'T KNOW" AND THAT IS WHY IT WORKS

three ways every answer can end, and only three:

→ sourced: it opened one file and named it
→ never stored: it tells you that instead of filling the gap
→ new: it writes a fresh file and adds one line to the map

that third one is what nobody builds. memory that grows without you managing it

the setup that makes it possible:
→ notes holds facts, one topic per file
→ people holds one card per person, role and how to reach them
→ projects holds status and the next step
→ root holds long term memory, lessons, and a log of why you decided things
→ the map sits at the bottom and loads before you ask anything

no apps. just markdown on disk

paste this into claude code in an empty folder:

"set up a second brain in this folder: 1. make three folders, notes, people, projects. 2. add root files for memory, lessons and a decision log. 3. write one map file that says where each kind of info lives. 4. one topic per file, lowercase with hyphens. 5. answer only from the file the map points to and name it. if nothing covers it, say so. show me the layout before creating anything"

the more you put in, the less it has to guess

[Bober_smart @Bober_smart]

https://t.co/WRvaPXgqIa

[Drerki @Drerki]

@kairon_flex same template as the bridge one, different set dressing. that is the part worth noticing

two girls, one helicopter, one fjord, and the same missing thing: no one hesitates before jumping

real people pause at the edge. the model never does

[Drerki @Drerki]

the bears are the whole trick. a crowd that does not react is exactly how real people behave when they think it is staged

watch the bounce physics instead. a body coming back up that high needs the trampoline to deform way more than it does

that is where these clips break first, not in the fall

[Drerki @Drerki]

the fan and coolant timing is the tell. 3,400 rpm and coolant still moving means it built a working state machine, not an animation

96% tensor load triggered by the chip exposure is a causal link. that is the model understanding what the part does, not just where it sits

second one from the same pipeline. that is repeatability, which matters more than the demo

[Drerki @Drerki]

the 92% drop is the part worth watching. limits shrinking that fast under moderate use means the cost per task went up, not that you used more

the two lines at the end of every run are the real trick. "what I reused" and "what I checked again" tells you exactly which file is missing

add that to your routine and the brief fixes itself

[Drerki @Drerki]

@z0renX one bot for everything is why it drifts. each new task pulls it toward a different average

the custom mode is the fix. lock the working style for that run and it stops re-reading the room

three bots with clear jobs beat one bot with a great prompt

[Drerki @Drerki]

@OmegaVault_app exactly. having your notes ready when you need them is the whole point

the setup you are showing saves hours of searching through old files

[Drerki @Drerki]

@krs1k exactly. guessing is now harder than admitting it is missing

[Drerki @Drerki]

CLAUDE FORGOT YOUR BANK PIN BECAUSE IT HAD NOWHERE TO LOOK

the fix is not a smarter model. it is one map file

a helix of plain text files twists around that map. every answer cites the exact file it came from

the setup:
→ one map file loads before you type anything
→ three folders wrap around it: notes, people, projects
→ one topic per file, lowercase with hyphens
→ memory and decisions stay pinned at the center

the rule that holds it together: answer from the file the map points to and cite it. not there? say so

no vector db, no Notion, no Obsidian. plain markdown on disk

claude stops guessing because looking is faster than inventing

paste this into an empty folder in claude code:

"build a second brain here: folders notes, people, projects. root files for memory, lessons, decisions. one map file listing where each kind of info lives. one topic per file, lowercase with hyphens. answer only from the file the map points to and cite it. if it is not there, say so. show me the plan first"

every file you add makes the helix worth more

[Bober_smart @Bober_smart]

https://t.co/WRvaPXgqIa

[Drerki @Drerki]

@Synthexfr exactly. guessing is now harder than admitting it is missing

[Drerki @Drerki]

@0xwhippa it is the router. everything else is just storage

[Drerki @Drerki]

@Stiingy7 start with passwords. that file pays back first

[Drerki @Drerki]

@0xLonkir the citation rule is the part that works. no source, no answer

[Drerki @Drerki]

writing a three.js model from a photo without any mesh input is the real jump

the 48 Hz wing stroke synced to a 3-stage reduction and the self-disassembly on a timed sequence means it is generating working kinematics, not just shapes

that is closer to CAD from a reference image than rendering

[Drerki @Drerki]

240 hz cursor update over bluetooth with zero latency means the sensor fusion and prediction model are running on device

the part apple charges for is not the hardware. it is the integration and the gesture library

claude rebuilt the gesture layer in 10 minutes. the hardware was already in your pocket

[Drerki @Drerki]

the 3 out of 12 filter is the part most copy bots skip

one wallet buying is noise. three wallets hitting the same coin in 10 minutes is coordination or information

the -8% hard stop is what kept the losers small. 11 winners at unknown R vs 5 losers at exactly -0.8R each is how you compound overnight

[Drerki @Drerki]

fixed price contracts on borrowed hardware with rising rates is a margin squeeze textbook

75% of revenue from three clients and 953bps on credit default swaps means one client delay turns into a liquidity event

the 16% annual default probability is the market pricing the refinancing risk, not the AI demand risk

[Drerki @Drerki]

the pre-flight check is the one piece most traders skip

calculating exact position size before the entry form opens is what separates a system from gambling

the tilt guard prompt after two losses is worth more than the whole scanner. that third revenge trade is where accounts die

[Drerki @Drerki]

@quz_ai the watermark one. invisible by default means it is already in every public grok api response you shipped

check x-watermark on your last 100 production calls and find out how many clients are serving watermarked text without knowing it

that is the issue, not the feature