Verdao

Dev / 23 / Focused on AI & Tech /

[Verdao @0xverdao]

YOU’RE USING 10% OF OPUS 5.5

every ls, git status and test run dumps a wall of noise straight into its context

one 30-minute session: ~118k tokens in, ~24k it actually needed

that’s why it forgets what you said 20 minutes ago and stalls mid-task

these 20 repos kill the noise before it ever reaches the model, 752k github stars between them

01 caveman (103.3k ⭐)
makes the agent answer terse, up to 65% fewer output tokens
▸ https://t.co/Y0ScdQjl8b

02 claude-mem (97.2k ⭐)
squeezes old sessions into compact memory
▸ https://t.co/b0jBs269K5

03 rtk (80.4k ⭐)
filters shell output before claude sees it, 60 to 90% less
▸ https://t.co/DHIfJt9LgJ

04 headroom (74.8k ⭐)
compresses tool output on the fly
▸ https://t.co/S3YkhM9Cc0

05 context7 (62.9k ⭐)
feeds only the fresh docs you need, not whole sites
▸ https://t.co/zg7JwLw9gl

06 litellm (53.8k ⭐)
one gateway for 100+ models
▸ https://t.co/2qO26ZhPG4

07 claude-code-router (37.3k ⭐)
sends each task to the right model
▸ https://t.co/QuZUgc6kuq

08 code-review-graph (32k ⭐)
a code graph instead of reading every file
▸ https://t.co/gHZf5y6VOk

09 repomix (28.6k ⭐)
packs a whole repo into one compressed file
▸ https://t.co/6Qv9D8w2hF

10 serena (27.9k ⭐)
pulls symbols, not full files
▸ https://t.co/I7z2pppNUo

11 context-mode (25.9k ⭐)
keeps raw tool output in a sandbox, out of context
▸ https://t.co/SzoxOTmJMW

12 toon (25.5k ⭐)
json for llms with far fewer tokens
▸ https://t.co/hpdMDf2p3R

13 SuperClaude_Framework (23.9k ⭐)
commands plus an ultra-compressed mode
▸ https://t.co/Yqfdbizrbh

14 ccusage (18.9k ⭐)
shows exactly where your tokens go
▸ https://t.co/IrUMBqYiVS

15 gitingest (15.9k ⭐)
any repo into a prompt-ready digest
▸ https://t.co/C4Ez6mSK8G

16 claude-context (12.6k ⭐)
semantic code search as an mcp
▸ https://t.co/cyPKQ6M9wz

17 Claude-Code-Usage-Monitor (8.6k ⭐)
live token burn monitor in your terminal
▸ https://t.co/EXCj9oVRSX

18 GPTCache (8.2k ⭐)
semantic cache, repeat questions skip the model
▸ https://t.co/aC0HiwVY3g

19 code2prompt (7.7k ⭐)
codebase into one prompt with a token count
▸ https://t.co/yh9I5tvEhy

20 LLMLingua (6.5k ⭐)
microsoft’s prompt compression, up to 20x
▸ https://t.co/OT8GiQFNkG

the bottleneck isn’t intelligence anymore, it’s what you let into the context window

[kiosa @thegreatest_sv]

https://t.co/6okVeHBq7p

[Verdao @0xverdao]

HOLY SH*T, OPUS 5.5 WITH THESE 10 REPOS FEELS ILLEGAL

out of the box it’s a very polite stranger: it doesn’t know you, can’t read the live web, chokes on your pdfs and forgets you by tomorrow

plug these in and it reads any site, eats any file, remembers you and clicks through your screen by itself

870k github stars between them, all free

01 firecrawl (190k ⭐)
turns any website into clean markdown claude can actually read
▸ https://t.co/Bmey3LGI0E

02 markitdown (188.9k ⭐)
microsoft’s converter, any pdf, docx, xlsx or pptx into markdown
▸ https://t.co/6rqE3Cu2Lb

03 browser-use (117.6k ⭐)
lets an agent click through the browser and do your boring tasks
▸ https://t.co/bL11q7kEos

04 MinerU (81.3k ⭐)
parses ugly pdfs, formulas and tables included, into llm-ready text
▸ https://t.co/SDl63apZvO

05 docling (68.4k ⭐)
document parser that keeps tables and layout intact
▸ https://t.co/zireht4iQc

06 mem0 (66.9k ⭐)
memory layer, so your assistant remembers you across chats
▸ https://t.co/pGweCNfvo6

07 fabric (44.2k ⭐)
a library of crowdsourced prompt patterns for real tasks
▸ https://t.co/d6G5OWiexe

08 UI-TARS-desktop (39.2k ⭐)
bytedance’s computer use agent that runs on your desktop
▸ https://t.co/1hnrfbG3zG

09 prompt-eng-interactive-tutorial (37.1k ⭐)
anthropic’s own prompting course, free, in notebooks
▸ https://t.co/v6Q9Eio2gB

10 khoj (36.4k ⭐)
self-hosted ai second brain over your notes and docs
▸ https://t.co/TD0yyvRvqs

the model is the same for everyone, your setup is the only edge you get

[Defileo🔮 @defileo]

https://t.co/N0MN5hBgHT

[Verdao @0xverdao]

80% OF WHAT OPUS 5.5 READS IS F*CKING TRASH

timestamps, permission bits, “counting objects”, 200 lines of passing tests

none of it helps, all of it eats your context

that’s why opus forgets what you told it 20 minutes ago and hits the limit mid-task

one 30 minute session went from 118k tokens to 24k without changing a single prompt

these 10 github repos cut the noise before it ever reaches the model

01 caveman (103.3k ⭐)
makes opus reply in terse caveman style, about 65% fewer output tokens
▸ https://t.co/Y0ScdQiNiD

02 claude-mem (97.2k ⭐)
compresses old sessions into memory instead of replaying everything
▸ https://t.co/b0jBs25BUx

03 rtk (80.4k ⭐)
filters shell output, 60 to 90% fewer tokens on everyday commands
▸ https://t.co/DHIfJt9drb

04 headroom (74.8k ⭐)
squeezes tool output, logs and files before the model sees them
▸ https://t.co/S3YkhM94ms

05 context-mode (25.9k ⭐)
keeps raw tool output in a sandbox, out of your context window
▸ https://t.co/SzoxOTmbXo

06 toon (25.5k ⭐)
a json format built for prompts, around 40% fewer tokens
▸ https://t.co/hpdMDf1Rej

07 ccusage (18.9k ⭐)
shows exactly where every token in your sessions went
▸ https://t.co/IrUMBqXL6k

08 claude-context (12.6k ⭐)
code search mcp so opus pulls matches instead of whole files
▸ https://t.co/cyPKQ6LBH1

09 Claude-Code-Usage-Monitor (8.6k ⭐)
live usage tracking that warns you before you hit the wall
▸ https://t.co/EXCj9oVk3p

10 LLMLingua (6.5k ⭐)
microsoft’s prompt compression, up to 20x smaller
▸ https://t.co/OT8GiQFfv8

git push alone goes from 15 noisy lines to “ok main”

[kiosa @thegreatest_sv]

https://t.co/6okVeHBq7p

[Verdao @0xverdao]

OPUS 5.5 IS WASTED ON YOU

most people open it, type a question, get a wall of text and close the tab

no project, no files, no rules, so every chat starts from zero

it has no idea who you are, so you’re talking to a very polite stranger

10 minutes of setup fixes that, and these 10 github repos do the heavy lifting

01 firecrawl (190k ⭐)
turns any website into clean data opus can actually read
▸ https://t.co/Bmey3LGI0E

02 markitdown (188.9k ⭐)
converts any file or office doc into markdown
▸ https://t.co/6rqE3Cu2Lb

03 docling (68.4k ⭐)
gets your pdfs ready for ai, tables and all
▸ https://t.co/zireht4iQc

04 fabric (44.2k ⭐)
crowdsourced prompts that solve real everyday tasks
▸ https://t.co/d6G5OWiexe

05 UI-TARS-desktop (39.2k ⭐)
an agent that clicks through your screen for you
▸ https://t.co/1hnrfbG3zG

06 prompt-eng-interactive-tutorial (37.1k ⭐)
anthropic’s own hands-on prompt tutorial
▸ https://t.co/v6Q9Eio2gB

07 khoj (36.4k ⭐)
an ai second brain that answers from your own files
▸ https://t.co/TD0yyvRvqs

08 gpt-researcher (29.5k ⭐)
deep research on any topic, every claim with a source
▸ https://t.co/kjPI6N3h2W

09 storm (27.5k ⭐)
researches a topic and writes a full report with citations
▸ https://t.co/zfCLwOLkh2

10 courses (22.3k ⭐)
anthropic’s free courses, from first prompt to tool use
▸ https://t.co/Afle6CMVPK

three files and seven rules, and it stops guessing who it’s talking to

[Defileo🔮 @defileo]

https://t.co/N0MN5hBgHT

[Verdao @0xverdao]

YOU’RE USING 10% OF OPUS 5.5

out of the box it has no skills, no rules and no team

it guesses your workflow, forgets your standards and works alone

1.5 million developers already starred the fix

these 10 github repos turn one model into a whole engineering team

01 superpowers (293.3k ⭐)
a skills framework that makes opus plan, test and ship like a senior
▸ https://t.co/BTCRswuFlZ

02 skills (284.1k ⭐)
real engineering skills straight from matt pocock’s own setup
▸ https://t.co/6rdTFKT2Zo

03 andrej-karpathy-skills (218.7k ⭐)
one CLAUDE.md file that kills the worst llm coding habits
▸ https://t.co/LQDQ4crYnu

04 claude-code (150.1k ⭐)
the agent that lives in your terminal and knows your codebase
▸ https://t.co/m6qzCsDjje

05 spec-kit (139k ⭐)
write the spec first, let opus build exactly that
▸ https://t.co/idvo8Xynyh

06 gstack (135.7k ⭐)
23 tools that give you a ceo, designer, qa and release manager
▸ https://t.co/mtgzjUHUfa

07 awesome-mcp-servers (95.2k ⭐)
the biggest list of mcp servers to plug into opus
▸ https://t.co/c0UaNAPB8g

08 servers (90k ⭐)
the official reference mcp servers
▸ https://t.co/FdpKpmwUru

09 ruflo (74.3k ⭐)
multi agent swarms with shared memory
▸ https://t.co/2FotgY6Ws0

10 claude-code-best-practice (66.1k ⭐)
the playbook from vibe coding to real agentic engineering
▸ https://t.co/Akfl6Z60ez

install the first three and opus stops guessing how you work

[kiosa @thegreatest_sv]

https://t.co/6okVeHBq7p

[Verdao @0xverdao]

THESE 10 REPOS SHOULDN’T BE PUBLIC

every ls, every git status, every test run dumps a wall of noise into its context

200 lines of a passing test suite when all it needed was “2 of 15 failed”

that noise eats your limits and makes opus forget what you said 20 minutes ago

these 10 github repos stop the leak before it starts

01 caveman (103.3k ⭐)
makes opus answer in terse replies, about 65% fewer output tokens
▸ https://t.co/Y0ScdQjl8b

02 claude-mem (97.2k ⭐)
compresses old sessions into memory instead of replaying them
▸ https://t.co/b0jBs269K5

03 rtk (80.4k ⭐)
filters shell output, 60 to 90% fewer tokens per command
▸ https://t.co/DHIfJt9LgJ

04 headroom (74.8k ⭐)
squeezes tool output, logs and files before the model reads them
▸ https://t.co/S3YkhM9Cc0

05 litellm (53.8k ⭐)
one gateway for 100+ model providers
▸ https://t.co/2qO26ZhPG4

06 claude-code-router (37.3k ⭐)
routes each task to the right model
▸ https://t.co/QuZUgc6kuq

07 code-review-graph (32k ⭐)
a local code graph so opus stops rereading whole files
▸ https://t.co/gHZf5y6VOk

08 repomix (28.6k ⭐)
packs a repo into one file, compress mode trims the tokens
▸ https://t.co/6Qv9D8w2hF

09 serena (27.9k ⭐)
semantic retrieval, opus pulls symbols instead of files
▸ https://t.co/I7z2pppNUo

10 context-mode (25.9k ⭐)
keeps raw tool output in a sandbox, out of your context
▸ https://t.co/SzoxOTmJMW

a 30 minute session went from 118k tokens to 24k with number 03 alone

[kiosa @thegreatest_sv]

https://t.co/6okVeHBq7p

[Verdao @0xverdao]

AFTER EFFECTS IS F*CKED

these 10 github repos turn opus 5.5 into a motion studio that renders while you sleep

01 anthropics/skills
the official skill format the whole studio runs on
▸ https://t.co/fbOry0mTaK

02 remotion-dev/skills
remotion rules for agents, no more guessing
▸ https://t.co/ILYZt5Z35l

03 greensock/gsap-skills
official skills so opus writes clean timelines
▸ https://t.co/c2FI12ToYa

04 heygen-com/hyperframes
write html, get an mp4
▸ https://t.co/ln8N1Ralal

05 remotion-dev/remotion
react components become video frames
▸ https://t.co/E2VI5nBzPc

06 greensock/GSAP
timelines you can seek to any frame
▸ https://t.co/2Jr6q28ueT

07 juliangarnier/anime
springs, staggers and svg morphs
▸ https://t.co/zxBVIcmPTF

08 microsoft/playwright
headless chromium screenshots every frame
▸ https://t.co/nHb6QDQvhK

09 FFmpeg/FFmpeg
turns frames into mp4 with motion blur
▸ https://t.co/wsiOV15ioq

10 librosa/librosa
builds the beat grid so every cut lands on the music
▸ https://t.co/iJpM2x6crO

587k stars combined, 0 seconds in after effects

[Carnage @0xCarnagee]

https://t.co/EUtvz78mOc

[Verdao @0xverdao]

ELON MUSK JUST PICKED CLAUDE OPUS 5.5 OVER HIS OWN MODEL FOR EVERY GROK BOT

on october 6 musk wrote that spacex will use the best back end model for any task, and the first name on his list was claude opus 5.5

the grok bot team confirmed it hours later, every bot now runs on opus 5.5

each bot has its own computer in the cloud, its own memory and its own logins, and they can spawn cursor agents and message each other in group chats

spacex kept the harness and anthropic got the brain

midjourney and suno were on the same list, and neither has a launch date yet

[Morlex @0xMorlex]

Grok Bot just got a serious brain upgrade

With Claude Opus 5.5 entering the stack, one Bot could run an entire AI team that researches, writes code, uses tools, verifies results, and delivers finished work

Here's how to build it in 10 steps: https://t.co/SysbtXuqoJ

[Verdao @0xverdao]

CHATGPT’S CO-INVENTOR WENT ROGUE
JEV + KIMI K3 SHOULDN’T EXIST

jev can’t write a single word, it only decides, and it does that in 70 to 500 ms

someone logged 1,100 calls from one research agent, and 1,003 of them were tiny questions like which step next or is this source good enough

now jev answers all of those, and the moment it’s less than 85% sure it hands the case to kimi k3 with the full context

100 fraud emails went through it, jev sorted them all in 1.42 seconds, kimi k3 took the 31 hard ones, and it got 96 right

the same layer took a claude session from almost 1M tokens down to 86K in one second

one model decides, the other one thinks, and your agent stops guessing

[starmex @starmexxx]

https://t.co/nJIv6r2vc7

[Verdao @0xverdao]

OPUS 5.5 JUST DELETED POLAND

i asked it for minecraft inside real street view and it went way too far

→ 15 TNT stacked on a real street
→ a gold and diamond tower
→ a giant TNT falling from the sky
→ craters blown into the actual road

1 html file, 0 libraries, 0 lines of code from me

which city do i blow up next

[jaivin @jaivinwylde]

I put the entire game of Minecraft inside twitter.. with WORKING multiplayer 💀 https://t.co/dpCLo0PvLp

[Verdao @0xverdao]

OPUS 5.5 BUILT ME A HOUSE IN POLAND, I ACCIDENTALLY NUKED IT...

built it block by block on my own real street, full minecraft style

one bag of tnt and the house is just gone

drop your city below, most liked one gets a house too, then gets blown up

[jaivin @jaivinwylde]

I put the entire game of Minecraft inside twitter.. with WORKING multiplayer 💀 https://t.co/dpCLo0PvLp

[Verdao @0xverdao]

THIS FEELS ILLEGAL TO WATCH: I RELEASED A SPIDER INSIDE OPUS 5.5 AND IT WENT FERAL

it broke into 10 data sources and nobody stopped it

codebase, docs, logs, database, inbox, it crawled every single one

→ 300 tasks ripped out
→ every source wrapped in silk
→ each task pushed through read, reason, verify
→ 25 failed the first time and got dragged back for a retry
→ 290 solved, 10 dead

44 seconds from the first jump to a finished web

the run is simulated, the workflow is real

this is how opus 5.5 actually eats a messy pile of work

watch it to the end, the last 5 seconds are the best part

[CyrilXBT @cyrilXBT]

https://t.co/SkcIaL0KmE

[Verdao @0xverdao]

this is the cheat code

i run a company where the only employee is me. a chief bot routes the work, the others do it, and i approve what matters

every grok bot gets its own cloud computer and keeps working after i close the laptop

>CHIEF - audits the business, picks the first 3 roles, routes every task
>ROLE - one narrow job per bot, named like a hire
>CHARTER - what it owns, what good looks like, where it stops
>HANDOFF - bots pass work to each other, i only get the judgment calls
>TAPE - no source, no number

most people pile everything onto one clever bot and it drowns in its own context

anything that sends, spends, publishes or deletes waits for my yes

[NO1ennn @N01ennn]

https://t.co/Js2rD31D6T

[Verdao @0xverdao]

THIS IS GETTING SCARY

run claude code's decision layer on your own laptop for $0

jev took the decision bill down to $3 a month
the real problem: every bash command, diff and .env line still left your machine, 600 times a night

two weeks later the open clones are here:

→ von: 1.5 GB, 18 ms on a mac, same endpoint as jev
→ kev: one base_url change, your jev hooks keep working
→ laya: 421M, #1 trending on hugging face

keep the narrow questions local:

→ secret in this edit? regex first, then von
→ which claude for this task? haiku, sonnet or opus
→ is this bash read-only? rules first

send only the fuzzy ones out

$3 to $0 is not the point
nothing leaves the machine is

[Asteri @Asteri_eth]

https://t.co/hJUKtndNOL

[Verdao @0xverdao]

MY CLAUDE LOOP SPENT $765 A MONTH ON YES-OR-NO QUESTIONS. NOW IT SPENDS $3

one overnight loop asks 600 small questions:

→ is this command safe
→ which file next
→ is the goal met

every one went to the same model that writes the code, about 4 cents and 5 seconds each

so i moved the deciding to jev:

→ safety gate: every bash command classified in 100 ms before it runs
→ stop hook: checks passes, jev says done at 0.93, the loop stops
→ triage cron: claude only wakes up for a reproducible bug

every decision hits the same line:

jev decides
claude only builds

same loop, 200 turns a night, $0.10 instead of $25.50

[Mr. Buzzoni @Mr_Buzzoni]

https://t.co/hIl5jyrAQz

[Verdao @0xverdao]

JEV RAN 10 BUILDS IN ONE AFTERNOON. EVERY ONE COSTS CENTS. ALL 10 ARE ALREADY SHIPPED BY SOMEONE

the idea is simple: jev does not write anything, you hand it state and a fixed set of questions, it hands back a label, a score or a probability

three rules run through all ten:

> ask everything in one call, more questions barely move the response time
> build the option list in code and always add "other"
> threshold on confidence, anything under 0.5 goes to a person or a bigger model

what people already shipped with it:

> inbox triage: 500 emails in seconds for 3.5 cents
> post scorer: 3,282 posts, 8 questions each, $0.1282
> rag rerank: 7 times faster than the hybrid it replaced
> tool-call gate: up to 18 times faster at p95
> context compaction: nearly 1M tokens down to 86K in about a second
> rows in sql: 129 rows judged for $0.0009
> ad library: 724 ads from 37 brands in 40 seconds for 9 cents
> citation checker: 1,709 judgments for under a cent
> pr triage: a merge read in about 200 ms

but the routing is what makes it work

> a confident answer goes straight to the queue
> an unsure one goes to a bigger model
> a risky one stops and asks a person

label 50 real examples by hand, run it in shadow mode for a day, switch it live when the log stops surprising you

[rody @0x_rody]

https://t.co/ZXF3ypv8za

[Verdao @0xverdao]

i asked jev to pick one billion-dollar startup and map how to rebuild its core product before i woke up

at 08:30 one brief was waiting in slack, with 2 decisions for me and 0 things sent without my approval

the whole night ran on one chain:
pick target → read public docs → spec → draft → audit → brief

5 lanes, each with one job and one hard limit:
→ jev picks the target, scores every startup, never writes a line
→ vance writes the spec and audits the result, never sends without sign-off
→ mara reads the public docs in the cloud browser, read and draft only
→ cole drafts 6 outreach briefs, zero automated emails
→ owen runs the pricing math, can't change a single record

astra plans and audits, sol does the volume at a fifth of the price

which billion-dollar startup should jev pick next?

[rody @0x_rody]

https://t.co/ZXF3ypv8za

[Verdao @0xverdao]

Kimi K3 + Jev make this AI stack look illegal

8 GitHub repos that take the cost out of every call

01 litellm
▸ https://t.co/Dgh4QuSWwO→ one gateway in front of 100+ models, spend tracked per call

02 RouteLLM
▸ https://t.co/SQlkulVktS→ easy requests go to the cheap model, hard ones to the strong one

03 GPTCache
▸ https://t.co/Ix0hz89RMl→ semantic cache, the same question never gets paid for twice

04 LLMLingua
▸ https://t.co/8ZToa4JlRG→ compresses the prompt before it is sent, fewer input tokens

05 vllm
▸ https://t.co/pbiYY1hM54→ serve open models on your own GPUs at high throughput

06 ollama
▸ https://t.co/hKbKItTEFy→ run models locally, zero cost per token

07 helicone
▸ https://t.co/FbFH9nkcez→ cost and latency logged for every request

08 promptfoo
▸ https://t.co/wl8rzkytVI→ evals that prove the cheaper setup still passes

the architecture:
route → cache → compress → host → measure → prove

i'd split the stack like this:

route:
litellm → RouteLLM

shrink:
GPTCache → LLMLingua

host:
vllm → ollama

prove:
helicone → promptfoo

Kimi K3 does the heavy work, the forks in between (which model, which tool, retry or stop) go to Jev at $0.042 per million input tokens

everyone posts the model and the benchmark

[Mr. Buzzoni @Mr_Buzzoni]

https://t.co/SxHnsCMK6g

[Verdao @0xverdao]

ONE PERSON. FIVE DOTS. EACH ONE OWNS ONE PART OF MY LIFE AND NOTHING ELSE

→ money: unpaid invoices, renewals inside 14 days, one brief every monday
→ calendar: guards deep work, flags any day over 3 meetings, briefs me sunday night
→ reading: 3 lines per link, flags when two saved pieces disagree
→ side project: where i stopped, why, and the one next action
→ admin: lays out the options, picks one, then stops

i wrote the line before the job
they draft, i approve, nothing sends alone

the exact prompts for all five are in the article below ⭣

[Verdao @0xverdao]

https://t.co/qmTk2eX5x6

[Verdao @0xverdao]

FIVE DOTS ORBIT MY LIFE ALL NIGHT AND NOTHING CROSSES THE LINE WITHOUT ME

one dot, one domain:

→ money caught $6,720 i forgot was owed to me
→ calendar flagged thursday, 4 meetings, deep work below 14h
→ reading cut 17 saved links down to 3 worth it
→ project knows i stopped at auth flow and the next action
→ admin found the insurance option $310 cheaper

every one of them drafts
none of them pays, sends, signs or publishes alone

0 crossed the line without me

the exact prompts for all five are in the article below ⭣

[Verdao @0xverdao]

https://t.co/qmTk2eX5x6