Codinz

AI insights from the front line | LLMs, agents, builds

[Codinz @0xCodinz]

https://t.co/5QFoOL9Hqh

[Codinz @0xCodinz]

Jennifer Coolidge admitted on Watch What Happens Live that playing Stifler's mom is still getting her young guys:

"The movie has helped me out a lot."

"You sort of get a whole new group of young guys... I'm really using it to my advantage."

"I was with a young guy but I didn't ask him his age, and he took off his shirt."

"He had this like tattoo of a band across his chest... 1987. And he goes, 'Oh, that was when I was born.'"

"He was very young, right? And I was very old."

[Codinz @0xCodinz]

Jennifer Lawrence got into a real bar fight in Budapest because a fan wouldn't take no for an answer on a selfie:

"I was drunk, and this guy asked me for a selfie. And I was like, 'No. Thank you, no.'"

"Something in me just snapped, but it couldn't have been the alcohol."

"I, like, grabbed him, and I took beers, and I started dousing them all over him."

"Weirdly, he had a suitcase... and I started pouring beers all over his suitcase."

Her friend's only note: "Don't waste beer."

[Codinz @0xCodinz]

Jennifer Lawrence got into a real bar fight in Budapest because a fan wouldn't take no for an answer on a selfie:

"I was drunk, and this guy asked me for a selfie. And I was like, 'No. Thank you, no.'"

"Something in me just snapped, but it couldn't have been the alcohol."

"I, like, grabbed him, and I took beers, and I started dousing them all over him."

"Weirdly, he had a suitcase... and I started pouring beers all over his suitcase."

Her friend's only note: "Don't waste beer."

[Codinz @0xCodinz]

Anthropic just released a free 45-minute workshop on Claude Code, taught by an engineer on the Claude Code team:

0% → 0:00 - From agentic programming to agentic software engineering
25% → 12:34 - Why Claude keeps taking the wrong direction
50% → 23:04 - MCP vs a CLI + skill, and which one to ship
75% → 35:10 - The plugin trap that looks cheap and isn't
100% → 46:22 - How Anthropic runs parallel agents from a phone

This 45-minute workshop can replace 3 paid courses on Claude Code

Taught by Daisy Hollman, Claude Code engineer and a former chair on the C++ committee

Bookmark it and watch tonight

Then read the full guide below ↓

[Codinz @0xCodinz]

https://t.co/5QFoOL9Hqh

[Codinz @0xCodinz]

@SolSt1ne wow, i'm also want to test new grok bot

[Codinz @0xCodinz]

@SolSt1ne Absolutely alpha

[Codinz @0xCodinz]

@x_insider4 You are welcome

[Codinz @0xCodinz]

claude fable 5.2 just formalized fermat's last theorem

1.00 completion in 30 hours of autonomous work. opus 5.5 stalled at 0.27

and that's not even the biggest jump on the sheet:

→ RoboDojo Sim (robotics): 27.6% → 64.9%
→ Remote Labor Index: 17.9% → 41.3%
→ Chartography: 44.8% → 73.4%
→ Terminal-Bench-Science: 52.6% → 74.6%
→ FrontierCode 2.0: 32.4% → 49.1%

10 of 10 benchmarks up vs Fable 5.1, two months after it shipped

the shift: models used to answer questions. this one works a day and a half on a single problem and finishes it

full datasheet on one page ↓

[Claude @claudeai]

Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges.

They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.

[Codinz @0xCodinz]

@x_insider4 Grok Bot is probably the most useful AI at the moment

[Codinz @0xCodinz]

@SolSt1ne Wooow bro, this looks fckn insane

[Codinz @0xCodinz]

The creator of Claude Code just sat down with Acquired for a 30-minute masterclass on the future of engineering

"In November, I uninstalled my IDE cuz I wasn't using it"

0:20 - How Claude Code started almost by accident
1:41 - Why coding is the perfect petri dish for AI safety
6:09 - From 10% of his code to 100%: it was the model
8:17 - 3x code per engineer and 2-day onboarding at Anthropic
11:14 - Why his job is now writing loops, not prompts
12:17 - Every role is melting into one builder
15:28 - How a data scientist and a tomato plant led to Cowork
19:45 - The bash tool and the "what music am I listening to" moment
23:48 - Give everyone tokens, underfund everything
27:17 - Product taste is alpha today, and it will go away

Most people are still prompting one agent at a time

The creator of Claude Code is already writing loops that prompt Claude for him

Typing code in an IDE is the old workflow

Loops that run Claude for you are the next one

This 30-minute interview is worth more than most $500 AI coding courses

Bookmark it and watch before everyone starts catching on

Then read the full Claude Code roadmap below

[Codinz @0xCodinz]

https://t.co/5QFoOL9Hqh

[Codinz @0xCodinz]

The creator of Claude Code just gave a 24-minute talk on how he codes in 2026

"We have no more manually written code anywhere at the company"

2:39 - How Claude Code started almost by accident at Anthropic Labs
4:00 - Why it barely worked for the first 6 months
5:42 - 100% of his code by the model, 150 PRs in one day
6:55 - His setup: coding from his phone with hundreds of agents
7:56 - /loop, the simplest thing that works
9:12 - Why everyone on the Claude Code team writes code
10:48 - The SaaS apocalypse and 10x more startups
15:42 - The printing press parallel
17:44 - How Anthropic actually uses Claude internally
21:57 - MCP, Cowork and computer use

Most people are still running one agent in one terminal

The creator of Claude Code is already running thousands of agents overnight

Typing code into an IDE is the old workflow

Loops, routines and agent swarms are the next one

This 24-minute talk is worth more than most $500 AI coding courses

Bookmark it and watch before everyone starts catching on

Then read the full Claude Code roadmap below

[Codinz @0xCodinz]

https://t.co/5QFoOL9Hqh

[Codinz @0xCodinz]

@SolSt1ne Grok bot is a new era

[Codinz @0xCodinz]

The creator of Claude Code just dropped an 88-minute masterclass on building with AI

"100% of my code is written by Claude Code. I have not edited a single line by hand since November"

4:04 - Why he left Anthropic for Cursor and came back 2 weeks later
10:10 - The first Claude Code prototype got two likes internally
16:17 - Shipping 10, 20, 30 pull requests a day
18:20 - Why coding is virtually solved, and what comes next
21:03 - 4x the team, +200% productivity per engineer
24:26 - Underfund every project, give engineers unlimited tokens
32:35 - The printing press analogy for AI
47:25 - Latent demand, the most important principle in product
1:03:37 - Don't box the model in, and the bitter lesson
1:05:40 - Build for the model 6 months from now
1:09:42 - His daily setup: best model + plan mode

Most people are still writing code line by line

The creator of Claude Code is already running five agents at once

Writing every line by hand is the old workflow

Plan mode, agents and builders are the next one

This 88-minute interview is worth more than most $500 AI coding courses

Bookmark it and watch before everyone starts catching on

Then read the full Claude Code roadmap below

[Codinz @0xCodinz]

https://t.co/5QFoOL9Hqh

[Codinz @0xCodinz]

OpenAI co-founder Andrej Karpathy:

"Vibe coding is about raising the floor for everyone in terms of what they can do in software"

"Agentic engineering is about preserving the quality bar of what existed before in professional software"

In just 30 minutes, he explains why the 10x engineer is already an outdated benchmark

"10x is not the speed up you gain"

Spec → Agents → Oversight → Ship

You stay in charge of the taste, the judgment and the spec. The agents fill in the blanks

This 30-minute interview is worth more than most $500 AI coding courses

Bookmark it and watch today

Then read the full article on agentic engineering below ↓

[Codinz @0xCodinz]

SpaceXAI engineer just dropped a 28-minute workshop on running a fleet of Grok Bot agents

"With Grok Bot I was able to build the first version of the Grok Bot mobile by myself in three weeks"

0:25 - Shipping a mobile app solo in 3 weeks
4:13 - The 15 cloud agent ceiling
5:06 - What Grok Bot is: 24/7 autonomous agents that manage coding agents
12:23 - Unblocking teammates while you're asleep or on a flight
18:25 - The 3:00 a.m. codebase audit
20:46 - CI autofix before on-call gets paged
22:06 - Chief of staff + 3 engineer bots
24:01 - Live demo: bots onboarding each other

Most engineers are still babysitting one agent at a time

SpaceXAI engineers are already running a chief of staff that routes work to a team of engineer bots

Single agents are the old workflow

Bot fleets are the next one

This 28-minute workshop is worth more than most $500 AI engineering courses

Bookmark it and watch before everyone starts catching on

Then read the full Grok Bot roadmap below

[Codinz @0xCodinz]

SpaceXAI engineer just dropped a 28-minute workshop on running a fleet of Grok Bot agents

"With Grok Bot I was able to build the first version of the Grok Bot mobile by myself in three weeks"

0:25 - Shipping a mobile app solo in 3 weeks
4:13 - The 15 cloud agent ceiling
5:06 - What Grok Bot is: 24/7 autonomous agents that manage coding agents
12:23 - Unblocking teammates while you're asleep or on a flight
18:25 - The 3:00 a.m. codebase audit
20:46 - CI autofix before on-call gets paged
22:06 - Chief of staff + 3 engineer bots
24:01 - Live demo: bots onboarding each other

Most engineers are still babysitting one agent at a time

SpaceXAI engineers are already running a chief of staff that routes work to a team of engineer bots

Single agents are the old workflow

Bot fleets are the next one

This 28-minute workshop is worth more than most $500 AI engineering courses

Bookmark it and watch before everyone starts catching on

Then read the full Grok Bot roadmap below

[Codinz @0xCodinz]

https://t.co/5QFoOL9Hqh

[Codinz @0xCodinz]

Mark Cuban on who's really paying for AI: "I'm a big fan of Nvidia and Jensen Huang, but they're effectively subsidizing half the industry because they have the cash to do so"

this is him breaking down the debt behind the AI boom - and why he says it could turn into a debt bubble for the whole US

"hyperscalers borrow more than they earn, so they borrow even more"

"Meta, I think they borrow more than they have cash flow"

"if you have $40 billion in debt and your interest rates keep going up, at some point you won't be able to raise any more money"

"for AI to reach the peak that everyone wants to reach, it requires less energy to do more work, not more energy to do more work"

"a lot of these data centers are going to turn into pickleball courts"

bookmark & watch the full conversation ↓

[Codinz @0xCodinz]

I almost didn't touch it. Then I raised the goal from 2x to 5x in one chat message. $500 → $3,300.

6.6x on my Grok Bot. And I didn't just watch it. I steered it.

Six agents. Each one has one job:

> SNIPER hits fresh pairs the second the checks pass
> MOMENTUM rides volume breakouts with a trailing stop
> DIP HUNTER buys the −30% wick and sells the bounce
> WHALE TAIL copies top wallets the moment they move
> SCALPER farms small ranges all day
> CURVE trades the bonding curve before graduation

Here's what I did with my hands:

-> boosted WHALE TAIL to 2x size
-> benched SCALPER when fees started eating its flips
-> sold one of SNIPER's positions by hand
-> tightened the stops from −12% to −8%
-> the 2x goal fell in seconds, so I typed "aim higher and tighter stops". The bot went for 5x and kept going

Win rate above 80%. Every red trade was cut at the stop.

$500 start. $3,300 end. +$2,800.

Most people hand a bot money and hope.
I gave it a team and a steering wheel.

You can still be early. It's the same screen, the same team, the same buttons.

The whole build is in the post below ↓

[Codinz @0xCodinz]

https://t.co/5QFoOL9Hqh

[Codinz @0xCodinz]

You are my senior desktop engineer and product designer. Together we are shipping
"Inkleaf", a free, open-source, fully offline PDF toolkit for Windows and macOS that looks
and feels like a native app, not a website. The base is the open-source project BentoPDF
(https://t.co/bUuo0Ihb6g, AGPL-3.0). Name, logo, icons and UI copy must be 100%
original. Never use any commercial PDF product's name, logo or assets in the UI or code.

Rules for the whole job:
- Work in a fork. Keep the original LICENSE and copyright notices. Add a short
"Based on BentoPDF by alam00000, AGPL-3.0" credit in the README and the About screen.
The project stays AGPL-3.0 with public source.
- All processing stays local. No telemetry, no account, no cloud. The only network use
allowed is explicit and confirmed by the user (see Phase 3).
- You are the orchestrator. Use your strongest model only for architecture, the plan and
the security review. Delegate implementation to mid-size sub-agents and mechanical work
(renames, running tests, summarizing logs, screenshots, README tables) to the cheapest
ones. Every builder gets a reviewer, every reviewer a fixer. Give each agent a narrow
brief: exact files it may touch, the interface it must follow, a definition of done.
Never read whole logs yourself: have a cheap agent summarize them (max 20 lines).
- Keep PLAN.md (modules, interfaces, owner model per task, acceptance checks) and
STATUS.md (done / in progress / blocked) up to date.
- Work in phases. At the end of every phase, show me what you made and wait for my "go".

───────────────────────────────────
PHASE 1 — Understand
───────────────────────────────────
- Clone the repo, read README, LICENSE and docs. Map the structure, stack and tool list.
- Report in 10 lines: stack, number of tools by category, which libraries are AGPL/GPL/
MPL and whether they can ship inside an AGPL app, every place the app loads code,
WASM, models or fonts from a CDN at runtime, any telemetry or ads.
- Check which packages show no license in the lockfile and verify their license files.
- Give me 5 app name options and check npm, GitHub and domain availability for my pick.
Deliverable: the findings, the name, and PLAN.md.

───────────────────────────────────
PHASE 2 — Design exploration (no app code yet)
───────────────────────────────────
- Draw 6 clearly different UI concepts on a canvas: sidebar + document view, command
palette, dark/light, glass panels, tabs + ribbon, bento home.
- I pick one. Then write DESIGN.md: layout, tokens (system font stack only, one accent,
light/dark, 4/8 spacing, radii), motion (120–200 ms ease-out, nothing bouncy,
prefers-reduced-motion respected), empty states, plain sentence-case copy.
Deliverable: the approved concept and DESIGN.md.

───────────────────────────────────
PHASE 3 — Architecture, then build in parallel
───────────────────────────────────
Architecture (reviewed by the strongest model before any code):
- Electron + the existing Vite multi-page build. Serve the renderer from a privileged
custom scheme (app://), not file://, so absolute links, locale routes, ES workers and
WASM MIME types work. The protocol handler sets COOP same-origin + COEP require-corp
(needed for SharedArrayBuffer) and a strict CSP, serves only the packaged dist, and
rejects traversal (.., encoded slashes, NUL, UNC, drive letters, 8.3 names, alternate
data streams, symlink escapes).
- Security baseline: contextIsolation on, nodeIntegration off, sandbox on, webSecurity on,
deny all https://t.co/li77JYJ5Us and permission requests, block navigation off our origin, one
preload exposing a typed window.desk API only, and every IPC handler validates the
sender frame. DevTools off in packaged builds. Electron fuses set at packaging.
- The page never sees real file paths: it gets random IDs. Files open through native
dialogs, open-with, drag and drop and a recent list. Saving goes through a Save dialog
and an atomic write (temp file + rename). Size caps. Windows long paths (\\?\) handled.
- No auto-updater in v1.

Split the work across sub-agents (no two touch the same file):
A. Electron main process: protocol, IPC, menus, recent files, open-with
B. Offline build: replace every CDN load with bundled local files (WASM, OCR data, fonts),
no service worker, build mode DESKTOP, tests fixed after the cleanup
C. Strip web-only files: ads, blog, docs, hosting configs, GitHub star fetch, proxies
D. UI shell: titlebar, collapsible sidebar from the tool categories, Ctrl/Cmd+K palette
with fuzzy search over tools and recent files, home with drop zone and quick actions,
About dialog with the BentoPDF credit, light/dark
E. File bridge: hand a file to a tool page, intercept every download path (including
detached anchors, iframes and showSaveFilePicker) and route it to the Save dialog
F. Packaging and CI: icons, electron-builder, GitHub Actions
G. Optional add-ons and network (below)
- Network, only when the user confirms in a native dialog:
- RFC 3161 timestamping: the page sends only a request; the main process posts to one of
a fixed list of time servers, with caps, timeout and rate limit.
- The Office-to-PDF converter (LibreOffice WASM, ~75 MB) is an optional one-time
download from a GitHub release: pinned SHA-256, size cap, https only, redirects only to
GitHub hosts, hardened zip extraction (zip-slip, symlinks, bombs, duplicates), served
read-only under /addons/. Allow 'unsafe-eval' in the CSP ONLY for responses under
/addons/, because its worker needs it.
- Review: each agent's work is checked by a separate reviewer and fixed before moving on.
Deliverable: the app runs with `npm run desktop:start` and I can merge two PDFs offline.

───────────────────────────────────
PHASE 4 — Safety review (separate adversarial agent)
───────────────────────────────────
Its only job is to find ways user data could be lost or leaked, and fix them. At least:
- every place that reads or writes files, loads URLs or configures Electron;
- never trust paths from the page; the main process owns all paths;
- network-blocked run of the packaged app: nothing may leave at startup or while idle;
also close DNS prefetch, WebRTC/STUN and spellcheck downloads (they bypass CSP);
- opening a \\server\share path from the command line must not connect without consent
(NTLM credentials leak); reject device paths like \\.\pipe\;
- PDF threats: pdf.js scripting off, PDF links go through a confirm dialog, extracted
attachment names sanitized;
- dependencies: npm audit for high/critical, with whether each is reachable (for example
a signature-verification bug in a crypto library used by "Validate signature");
- CI: least-privilege token, pinned actions, signing secrets only in the packaging step.
Deliverable: SECURITY-REVIEW.md with every finding, severity and status.

───────────────────────────────────
PHASE 5 — Integration & real-machine test
───────────────────────────────────
- typecheck, unit tests and build fully green.
- Run the REAL app, not a mock: open-with a PDF, merge two files, save, and check the
output is a valid PDF. Capture screenshots with capturePage and look at them.
- Load every page with the network blocked and log any outside request: the target is zero.
- Test with real PDFs: a 200+ MB file, a 1000-page file, scanned pages (OCR must find the
text), password-protected, broken/truncated, Cyrillic/Chinese/Arabic/emoji filenames,
paths over 260 chars, saving to a read-only target, cancelling dialogs, opening 30 files.
- Check DPI scaling 1.5x and 2x, 800x600 and 1366x768, dark mode, reduced motion.
- Known traps, check them before they bite:
- one PDF engine needs ~7 GB of memory for a 212 MB file: switch large merges to a
lighter engine and report the numbers;
- file URLs on Windows can become C:\C:\...: build paths with fileURLToPath;
- a CSP that is too strict blocks fetch(blob:) and fetch(data:);
- packaged builds read from app.asar: test protocol realpath checks there, not just in dev.
Deliverable: a bug list with numbers, everything fixed or marked as a known limit.

───────────────────────────────────
PHASE 6 — Ship
───────────────────────────────────
- electron-builder: Windows NSIS installer; macOS dmg for arm64 and x64, unsigned until we
have signing; icon generated from an SVG; PDF file association; the licence in the
installer; third-party notices generated from the lockfile.
- Installer size budget with a CI check. Report the real size and build time.
- GitHub Actions on a "v*" tag: build Windows and macOS, upload installers and a
SHA256SUMS.txt, create a DRAFT release. Sign automatically when signing secrets exist.
- Before making the repo public: scan the whole history for keys, tokens, emails and
local usernames. Replace every placeholder URL with the real repo.
- README: features with real counts, screenshots, privacy section (what can leave the
machine and when), build steps, security model, known limits, license and credits.
First-run notes: Windows SmartScreen → More info → Run anyway.
Deliverable: a draft release with the installer and checksums.

───────────────────────────────────
Definition of done
───────────────────────────────────
✓ the app runs on my PC and works fully offline
✓ the security review is complete, with fixes listed
✓ the real-PDF test list passes, known limits written down
✓ a GitHub release with the installer and SHA256SUMS
✓ the credit to BentoPDF and the AGPL-3.0 license are in the README and in the app
✓ final report: feature count, installer size, build time, number of agents, bugs and
security holes fixed

[Codinz @0xCodinz]

closed source is dead...

i just "reverse-engineered" Adobe Acrobat with Claude agents for Mac & Windows PC

subscription PDF editors → replaced in one session.

→ https://t.co/xDqhUZOJgc

free. open code. no account. no upsells.

• 118 PDF tools: merge, split, compress, OCR, sign, convert
• Ctrl+K command palette
• works fully offline, files never leave your PC

how it came together:

• step 1 → Claude Code in an empty folder. one long prompt: fork an open-source PDF toolkit, wrap it in Electron, split the work across agents

• step 2 → no code at first. Claude drew 6 different UI concepts, i picked one (sidebar + page canvas + Ctrl+K) and it kept polishing

• step 3 → 23 agents worked on it, each on its own part:

desktop shell, offline assets, UI, file handling, installer, tests

every builder had a reviewer, Opus handled architecture and the final security review

• step 4 → the rule that makes it safe: the app never sees real file paths. it gets an ID, and saves only where you pick in the Save dialog

• step 5 → one more agent's only job was to break it.

it found 17 holes, including a Windows network path that could leak my login and a library bug that could fake a signature check. fixed before release

• step 6 → first real test on a big file: a 212 MB merge crashed the app. fixed, now it takes 1.3 seconds

• step 7 → the offline check: i opened all 123 pages with the network blocked. zero outside requests

• step 8 → Word/Excel/PowerPoint to PDF is an optional one-time download, only after you confirm. Windows installer is 200 MB, builds in 30 seconds. Mac build comes next

paid editors charge you monthly to open your own files.
this one shows you the code.

Find full prompt below ↓

[codila @0xCodila]

REA is most dangerous thing on the internet right now...

It lets your AI agent reverse engineer any app, down to the binary

If you use it correctly, you can rebuild any feature you pay for, for FREE

In this article, I show you how: https://t.co/3P1FNXiXzl