Zyvqor

building and visualizing AI agent systems. I test the viral AI claims so you don't have to.

[Zyvqor @Zyvqor]

@X1Kafadzhy You don't let the router decide on security at all. Anything touching auth, secrets or permissions gets routed to security by a hard rule, before the score even matters. Confidence is for the boring calls.

[Zyvqor @Zyvqor]

Your AI reviews a typo fix and an auth change the same way. That's the bug.

Most setups throw every diff at the same review. A full deep review on a README typo. The same quick pass on a PR that touches secrets.

The fix: before any review runs, make one small call that decides how to review.

What goes in:
diff size
paths touched
test delta
number of pushes

What comes out is one choice from a list you define:
security
tests
quick look
deep review
ask a human

Plus a confidence score.

The threshold is the part that matters. In the video:
auth + secrets change → security, 0.94
README typo → quick look, 0.97
billing refactor with zero test changes → the top answer was only 0.37, so it goes to a human

The model doesn't get to guess on the risky stuff. Below 0.80, a person decides.

How to set it up:

1. Send a summary of the diff, not the whole diff. Small input, fast answer.

2. Keep the options fixed. A pick from 5 labels, not an essay.

3. Use a small fast model. Haiku can handle it, and decision models like Jev from TypeSafe are built for exactly this kind of call (still on a waitlist).

4. Pick a threshold and send everything below it to a human.

5. Log every call with its verdict. After a week you'll see where the router gets it wrong.

Run it in CI before review, and the expensive review only happens where it's needed.

One call per diff. The verdict moves with it.

[rody @0x_rody]

https://t.co/JDUqNIHixU

[Zyvqor @Zyvqor]

@X1Kafadzhy @nvidia @NVIDIAAI Game studios spend months on mocap just to make a character climb stairs without looking drunk. This does it in 2 ms.

[Vyrqo_ @vyrqo_]

24 free GitHub repos that turn Claude Code into a full video studio.

The video above was made with code. No After Effects, no timeline app. Just repos and a terminal.

Most people run Claude Code with nothing on it. No memory, no taste, every token burned on raw logs. Here's the toolkit.

PART 1. MOTION

1. Three.js: 3D scenes
https://t.co/5aNRcLyrUS

2. faster-whisper: auto captions from audio
https://t.co/OlqAp7oGDO

3. Remotion: videos made with React
https://t.co/IhgwRSW1wl

4. tsParticles: particle effects
https://t.co/CX8rfrxbuC

5. Tone.js: synths and sound fx
https://t.co/1OzHeFQwvu

6. PixiJS: fast 2D graphics on the GPU
https://t.co/fhYCx3BHF1

7. Chatterbox: open-source voiceovers
https://t.co/uwCVwqmTte

8. Matter.js: 2D physics
https://t.co/rRnUBh1cNj

9. HyperFrames: turns HTML into MP4, has a Claude Code plugin
https://t.co/1UWtIcUHkV

10. GL Transitions: shader transitions
https://t.co/Uq2XqJXsd2

11. Anime.js: timeline animation
https://t.co/TKcwNbH1lr

12. GSAP: the animation library pros use, now free
https://t.co/TZY5Phc2eu

13. FFmpeg: encode and export everything
https://t.co/6eDF69xOoK

14. Playwright: drives a real browser, so the agent can screenshot and render frames
https://t.co/xuCQo62RxB

PART 2. CLAUDE CODE

15. claude-mem: remembers what your agent did last session and brings it back
https://t.co/TT1xBNkIhE

16. context-mode: keeps raw tool output out of your context window
https://t.co/FTe4reSxWu

17. rtk: compresses command output before your agent reads it
https://t.co/oLmfuQnjiO

18. planning-with-files: the plan lives on disk, so /clear doesn't wipe it
https://t.co/dBfA4owMBk

19. agent-skills: Addy Osmani's workflow from spec to ship
https://t.co/uW6QDVlyvM

20. UI UX Pro Max: a design skill so your UI doesn't scream "AI made this"
https://t.co/Ua7wh6Wnqf

21. Everything Claude Code: plan, test, review, verify, remember in one plugin
https://t.co/LWKhAJHYa1

22. Agent-Reach: lets your agent read platforms it normally can't reach
https://t.co/wyF0mIZgMJ

23. Ruflo: runs a crew of agents across Claude Code and Codex
https://t.co/hoiI98shyU

24. career-ops: paste a job post, it checks the fit and drafts your answers. You still press send.
https://t.co/ThT3t8Y2Q8

Quick tip: don't install all of this.

Every plugin eats context before you type a word. Pick 3: one for memory, one for tokens, one for the job you actually do.

Run the same task before and after. If it's not better or cheaper, delete it.

[Zyvqor @Zyvqor]

@whemohere Data brokers are the only business where you're the product and you still have to ask to be removed.

[Zyvqor @Zyvqor]

@0xTimQ Ants turned $5 into $19K. My ants just ate my sugar.

[Zyvqor @Zyvqor]

@moatmemo Source: BBC Newsnight, 2000. Which company today looks "famously unprofitable" for the same reason?

[Zyvqor @Zyvqor]

@X1Kafadzhy Revenue up 13x, but sales and marketing went from $120M to $3.2B in the same period. Growth isn't free.

[Zyvqor @Zyvqor]

@DenisDS8 I think that's the point of letting them message each other. Ideally only the architect session talks to you and the rest report to it, so you're watching one window, not five.

[Zyvqor @Zyvqor]

One prompt. 5 Claude Code sessions. 9 messages between them. Zero copy-paste.

Most people run one Claude Code window and make it do everything. Planning, coding, migrations, tests. All in one context that gets messier every hour.

You can split it. Claude Code sessions on the same machine can now message each other.

The setup from the video:
opus: plans and reviews
sonnet: splits the work and routes it
three haiku sessions: frontend, api, db

You give the task once. Opus sends the plan to sonnet, sonnet hands each haiku its piece. When the db session alters the table, it tells api directly. Api ships the endpoint and tells frontend to regen its types. Opus reviews at the end and says ship it.

How to try it:

1. Update Claude Code. You need v2.1.224+ on Mac and Linux, v2.1.234+ on native Windows. After that it's on by default.

2. Open a few terminals and start each with its own model: claude --model opus, claude --model haiku. Give them names with --name so they're easy to call.

3. Type /list-agents to see which sessions can reach each other.

4. Then just ask: "let @api know the schema migration finished." Claude writes and sends the message itself.

Worth knowing:

A message is plain text, never the other session's history. Every window keeps its own clean context.

On one machine it travels over a local socket, not through Anthropic's servers.

A message from another session can't approve permissions or change your config. Each terminal still asks you.

For a background worker that should take messages without you: --settings '{"crossSessionInbound":"accept"}'

My favorite small one: "tell me when the migration session finishes." No more checking tabs every two minutes.

Bookmark this for the day one context window stops being enough.

[mirku @mirku21]

https://t.co/PaAzxlyrsH

[Zyvqor @Zyvqor]

@pantherbllack Which models do you actually get at these limits? The daily number means a lot less if it's only the smaller ones.

[Zyvqor @Zyvqor]

@0x_kern The time number is the underrated one. 105 s to 85 s is about 19% faster per answer. In an agent loop that runs the model 30 times in a row, that's the difference you actually feel.

[Moat Memo @moatmemo]

In 1974, Elvis wanted to record Dolly Parton's "I Will Always Love You." His manager's condition: half the publishing.

She said no and "cried all night."

In 1992, Whitney Houston's version spent 14 weeks at No. 1. The publisher was Dolly.

She tells Howard Stern why below.

[Zyvqor @Zyvqor]

@moatmemo Wilder detail: she wrote "I Will Always Love You" on the same day she wrote "Jolene." Two of the most valuable copyrights in country music, one sitting, and she kept 100% of both.

[Moat Memo @moatmemo]

HE BLAMED ONE BILLIONAIRE FOR CRASHING HIS COUNTRY AND SAID IT AT THE IMF. Mahathir, Hong Kong, 1997: currency trading is "totally immoral."

Soros called him "a menace to his own country."

A year later Malaysia pegged the ringgit at 3.80. The peg held until 2005.

[Zyvqor @Zyvqor]

@qwinsi0x The honest part of that image is the loop: research, code, backtest, gate, deploy. The gate is where small teams die. A self-improving loop finds hundreds of strategies that only work on the data it was tuned on. Without a hard out-of-sample gate it just loses money faster.

[Zyvqor @Zyvqor]

@vyrqo_ Curious what the agent does with a gap the spec doesn't cover. Does it flag it as "no reason" or make one up? The second one would be worse than the original problem.

[Zyvqor @Zyvqor]

YOUR PHONE NUMBER IS PROBABLY ON SITES YOU NEVER SIGNED UP FOR, AND AN AGENT CAN FILE MOST OF THE REMOVALS WITHOUT YOU

Most of that work is boring and repeatable: find the listing, check it's really you, fill a form with 3 fields. The risky part is small: an email link, an SMS code, an ID upload. So the skill splits the two before it starts.

the mission:

→ search people-search and phone-lookup sites for your number
→ confirm each listing matches you before touching it
→ fill the removal form with only your name, number and the listing URL
→ stop at any email link, SMS code, captcha, ID upload or payment
→ recheck every 30 days and refile if a listing comes back

Most people hand an agent the whole job or none of it. The fix is boring: write the boundaries before the first run. Anything the agent can undo, it finishes. Anything that proves who you are or costs money waits for you.

the same split works for a small team:

→ Lead owns the mission and the boundaries
→ Search returns every listing with its URL
→ Filer fills the forms and never sees your ID
→ Checker reruns the search and flags anything that came back

the prompt ↓

"Run a removal mission for my phone number [your number].
1. Search people-search and phone-lookup sites. Return each listing with its URL.
2. Show me the matches and wait until I confirm which ones are mine.
3. For each confirmed listing, fill the removal form with only my name, number and the listing URL.
4. Stop and hand me anything that needs an email link, SMS code, captcha, ID or payment.
5. Save the list and recheck it every 30 days."

no boundary, no autopilot.

where it gets hard:
→ a site can relist your number weeks later, so the monthly recheck is the whole point
→ the confirm step slows the first run, and it is what stops the agent from removing someone else's listing
→ captchas stay with you. an agent should never solve them for you

The full agent guide is in the article below.

[Sprytix @Sprytixl]

https://t.co/aRj8erlMO9

[Zyvqor @Zyvqor]

@Bober_smart My AI made a plan too. Still waiting for the clients.

[Zyvqor @Zyvqor]

CLAUDE CODE CAN'T SEE INSIDE YOUR IPHONE. IT CAN STILL CHECK MORE THAN YOU THINK.

My feed this week is phones where Claude Code "found the spyware", two of those posts passed 600K views. So I mapped what it can actually read on a phone, layer by layer, from the public docs.

android, over adb from your laptop:
→ pm list packages -3 lists every app you installed
→ dumpsys package shows camera, mic, location and SMS access per app
→ enabled_accessibility_services shows which apps can read your screen
→ enabled_notification_listeners shows which apps read every message preview
→ dumpsys device_policy shows apps that made themselves hard to remove
every one of these only reads. nothing gets installed on the phone and nothing gets deleted.

iphone, no adb:
→ ideviceinstaller -l lists your apps after you tap Trust
→ MVT checks an encrypted backup against published spyware indicators
→ profiles and signed-in devices you check by hand, in VPN & Device Management and Safety Check
→ data inside other apps stays sandboxed on both

no match is not a clean phone.

MVT's own docs say public indicators can't prove a device is clean. And if someone may be watching you, removing the app tells them. Make a plan first, then clean.

[starmex @starmexxx]

https://t.co/behHzfWQ7D