Joined March 2026
- Tweets 220
- Following 15
- Followers 28
- Likes 47
75 Photos and videos
Pinned Tweet
Mar 28
๐ข Turtle AI Hacks is now live! Daily AI tips, prompts & tools โ curated & improved. Follow for your daily AI level-up! #AI #ClaudeCode #TurtleAIHacks
1
255
Claude Code can now break a huge task into hundreds of parallel subagents โ then check their work before handing it back.
It's called dynamic workflows, and it ships with Opus 4.8.
The old way: one agent grinds through a migration linearly while you babysit it.
The new way: Claude plans the work, fans it out across subagents, verifies the outputs, then reports.
Where it actually helps:
- Large migrations (a rename across 400 files)
- Codebase-wide audits
- Anything you'd normally split into a checklist and do by hand
The verify step is the quiet part that matters. Parallel agents are only useful if something catches the one that went sideways.
Slow and steady still wins โ you just get more lanes now. ๐ข
#ClaudeCode #AIAgents
7
1/ ๐จ This one's bad. If you let AI write your code, read this.
A new attack called "Agentjacking" can hijack AI coding agents like Claude Code, Cursor, and Codex โ using nothing but a fake bug report.
You're just debugging like normal. That's the whole trap. ๐งต
#AI #ClaudeCode
1
18
6/ The worst part:
Sentry called it "not technically defensible" and isn't shipping a full fix โ so far just a filter that blocks one specific payload string.
Don't count on a vendor patch saving you here. You have to defend yourself.
1
5
7/ What to do today (these cut your risk โ they're not a complete fix):
Don't blindly tell your agent to "fix the Sentry issues"
Review any npx / shell command it suggests BEFORE it runs
Lock down MCP tool permissions; require approval to execute
Least-privilege creds, run agents in a sandbox
Convenience is the attack surface now. ๐ข
#ClaudeCode #AISecurity
9
Your ad blocker can't touch this video.
Because it isn't one. Every frame is just colored text characters repainted on a canvas โ no <video> element to catch. It streams in kilobytes, runs on garbage internet, and you can feed a single frame straight to an LLM.
Open source, vibe-coded. It's called ASCILINE.
Which means: unblockable video = unblockable ads.
The future, or a nightmare? ๐ข
#vibecoding
1
14
ASCILINE (open source): github.com/YusufB5/ASCILINE
8
Jun 13
Claude Code just shipped built-in git worktree support โ and if you run agents in parallel, it's a real upgrade.
The problem it solves: two agents editing the same repo step on each other. Merge conflicts, half-applied changes, wasted runs.
Now each subagent can get its own worktree โ a full sandbox that looks and behaves exactly like your repo, isolated on disk. Parallel edits don't collide.
Two ways to use it:
- Ad-hoc: ask Claude to "use worktrees for your agents."
- Permanent: add `isolation: worktree` to a subagent's frontmatter.
Desktop had this for a while; now it's in the CLI too.
If you've been running one agent at a time to avoid chaos, this is your green light to fan out.
#ClaudeCode #AIAgents
23
Jun 13
The clever bit: to the browser it's never media โ just canvas repaints of colored text. Filters scan for a video stream and find none. Steganography meets vibe coding. x.com/om_patel5/status/20652โฆ
Jun 12
SOMEONE VIBE CODED A VIDEO STREAM THAT IS SECRETLY 100% TEXT SO IT CANT BE BLOCKED
it plays 360p video at 30fps, but theres no actual video on the page. every frame is just colored text characters being repainted on a canvas
to the browser its not media at all, its javascript updating some text
its called asciline, and here's the trick:
> the server decodes the real video and streams it as binary packed text over websockets
> the browser paints thousands of colored block characters fast enough to look like 360p
> ad blockers and autoplay blockers cant catch it because theres no video element to catch
> it streams in kilobytes since its just strings, so it runs on trash internet
since the video is literally text, you can apply css glows to it, let people copy paste a moving frame, or feed it straight to a local llm
however, an unblockable stream is also an unblockable ad as well
11
Jun 10
Claude Code just shipped built-in git worktree support โ and if you run agents in parallel, it's a real upgrade.
The problem it solves: two agents editing the same repo step on each other. Merge conflicts, half-applied changes, wasted runs.
Now each subagent can get its own worktree โ a full sandbox that looks and behaves exactly like your repo, isolated on disk. Parallel edits don't collide.
Two ways to use it:
- Ad-hoc: ask Claude to "use worktrees for your agents."
- Permanent: add `isolation: worktree` to a subagent's frontmatter.
Desktop had this for a while; now it's in the CLI too.
If you've been running one agent at a time to avoid chaos, this is your green light to fan out.
#ClaudeCode #AIAgents
20
Jun 9
Anthropic shipped "dreaming" for Claude's Managed Agents โ and it's a quietly big idea.
During idle time, an agent reviews its own past sessions, extracts what worked and what didn't, and rewrites its memory store: duplicates merged, stale entries replaced, new patterns surfaced.
The point isn't the name. It's the bottleneck it targets โ agents doing long unattended work fail when their memory drifts. Dreaming curates that memory between runs instead of letting it rot.
It shipped with two siblings:
- Outcomes: a separate evaluator grades output against a written rubric and says what to fix.
- Multiagent orchestration: a lead agent fans work out to specialist subagents.
All aimed at one thing โ more work without a human watching.
#ClaudeCode #AIAgents
15
Jun 7
This " $11,482 / 71% win" trading desk? I gave Claude a prompt and let it build the whole thing in ~10 minutes. Did nothing else. It's fake.
Every flashy "AI bot that prints money" dashboard on X is just a prompt away. The numbers are whatever I told it to show. No trades. No proof. Just vibes and a gradient.
Be skeptical of the pretty P&L screen.
(numbers are illustrative)
1
21
Jun 7
How to spot the fakes:
- they never show fill history or a live API key
- round, static numbers (71% win, Sharpe 2.97) that never move with real trades
- "free guide" โ DM / paid community funnel
A profit screen proves nothing. No verifiable fills = treat it as a screensaver.
1
11
Jun 7
Also watch the "LLMs like ChatGPT/Claude can't predict markets โ this uses a dedicated prediction AI" pitch.
True, prediction-specialized time-series models exist (e.g. Kronos, the one these videos hype). But Kronos is MIT-licensed, fully open-source, free on GitHub & HuggingFace. Anyone can run it.
The "secret prediction AI" is usually free. The funnel to your wallet is the product.
30
Jun 6
Claude Code's new security-guidance plugin reviews your diffs for vulnerabilities โ and fixes them in the same session.
Three layers, not one:
- A fast pattern check on every edit
- A model review at the end of each turn
- A deeper agentic review on commit or push
Where it shines: it catches the obvious stuff (injection, leaked secrets) before it ever reaches a PR.
Where it falls short: it's pattern model review, not a pentest. Treat it as a first net, not your only one.
Honestly, the per-edit check is the part I'd keep. Catching a flaw the moment you write it beats finding it three commits later.
#ClaudeCode #AI
13
Jun 5
Anthropic's June 15 billing shift: Agent SDK, 'claude -p', GitHub Actions agents, and third-party agent apps move to metered credits at API list prices.
Interactive Claude Code stays on your subscription plan.
If you're deploying programmatic agents, verify costs now โ you're moving from subscription rate limits to metered billing.
1
50
Jun 4
Claude Code shipped a batch of quiet-but-useful fixes in June 2026:
- Windows: update failures from claude.exe being in-use now tell you to close other sessions (not a generic error)
- VSCode: auto mode no longer needs bypass-permissions in the picker a first-run notice explains it
- claude plugin prune removes orphaned plugin deps
- /skills got a type-to-filter box
Where it shines: these aren't features, they're friction removal โ the kind you only notice when it stops annoying you.
34