🤖 I build fun stuff with AI 🌐 Building https://xitter.twitterapi.xyz/t.co/cxu79YE9PW - 1 app for ALL your apps 🐶 Dog dad
Budapest, Hungary
Joined May 2025
- Tweets929
- Following1K
- Followers935
- Likes783
Pinned Tweet
Claude Code can delete your whole repo in one command..😳
The .git, commits, branchs, all gone
The 30 sec fix in .claude/settings.json:
"permissions": {
"deny": ["Bash(rm -rf:*)", "Bash(git push --force:*)"]
}
Now those get blocked even if it tries. Still push your work often tho!
Anthropic lets Claude "dream": go back over past work, fix its mistakes and rewrite its notes
you can do a mini version every friday for example, paste this into Claude code:
"look at this week's commits and diffs in this repo. update claude.md: add rules you learned and why, remove rules that are no longer true and flag anything youre unsure about. show me the diff before you save"
2
1
293
Claude Code: Fable 5.2 is here for your longest and toughest coding tasks!
also Claude Code: it draws down usage faster than Opus 5.5
ok so the toughest tasks, maybe twice a week..
1
1
73
202 agents, 185 skills and 105 commands for Claude Code all in one repo! (42k stars)
wshobson/agents
> split into 94 plugins so you only install what you need
> written once and adapted for codex, cursor, copilot and more
→ checked with evals not just a pile of prompts
grab 1 plugin that matches your stack, not all 94:)
7
did you know you can tag @claude in a Github PR comment and it works on it right inside the PR?
setup is one command inside Claude Code:
/install-github-app
it installs the app, adds the secret and opens the workflow PR for you
only works for github repos tho, not gitlab or bitbucket
9
1 in 4 AI requests on Vercel now go to models that dont write a single word
theyre called decision models
instead of generating text, they classify, score or pick between options (at about 1/70th the cost per request!!)
Google just launched the Gemini decisions API yesterday, up to 1000 questions in 1 call
if your app sends a full LLM call just to get a yes/no, spam or not spam or a category, thats the 1st thing to swap
7
Alibaba just dropped Qwen4 Max with a 1 million token context window!
but
> its only live for 2 weeks then gone on nov 25
> no benchmarks, weights or per token price
> the plan doesnt even let you build it into your own app
a model you can rent for 15 days and cant ship is a marketing test imo, not something to build on
1
1
103
2 things in the Claude code /model menu you shouldnt overlook:
1. the left and right arrows change the effort level (low for renames and small fixes, high for the hard stuff)
2. press S instead of enter to switch only this session so your default stays the same
running everything on high effort is a super surefire way to burn thru your limit🥲
11
a YC startup just launched a phone where the AI uses your apps for you!
its a Pixel 10 pro running their own Android, $1,299 or free with an iPhone trade in
the AI works across every app, reads your texts, emails and notifications and handles all the boring stuff you can think of
founder says they switched 10 people from iphone and none wanted theirs back lol
are you ok with one app reading every message you get?🤔
1
1
80
this is actually WILD
Vercel routes 100 trillion+ tokens a month thru its AI gateway and the new numbers say:
> more than half of language model requests go to models less than 2 months old
> 3 in 4 teams drop a model within a month of its peak
> new teams picked Anthropic nearly 2 to 1 over OpenAi
> the biggest spenders use a median of 29 different models
the model youre using today probably isnt the one you'll use in 2 months. dont hardcode it, keep it in one config line
11
Nvidia's board reportedly met last night to decide whether to fire Jensen Huang (who built the company over 33 years)
shares fell 8% after hours and Nvidia has lost $450B+ in value since nov 8..
most of the AI industry runs on his chips. ONE boardroom fight can ripple into prices, limits and launch dates for every builder
12
me: fix this one bug
Claude: *edits 14 files*
me: what did you change?
Claude: just a few small improvements
...
anyone else?🥲
2
2
99
Anthropic just published what Claude's best memory notes look like! (and most claude.md files do the opposite)
in their tests the models that actually learned wrote notes with 3 parts:
> the rule
> how they knew the rule was true
> the trap that made them write it
most claude.md files are just rules with no reason so the agent cant tell which ones still matter..
so next time you add a line, add the why
"use pnpm" becomes "use pnpm, npm breaks the lockfile in this repo"
14
paste this into Claude Code before it builds anything big:
"before you write any code:
1. tell me in 2 sentences what you think i asked for
2. list every file you plan to touch and why
3. tell me what could break and how we'd notice
4. ask me anything you're unsure about then wait
no edits until i say go"
it catches the wrong assumption before it turns into dozens of files of wrong code!
21
166 ready made subagents for Claude Code sorted into 10 categories in 1 repo:
VoltAgent/awesome-claude-code-subagents (26k stars)
> backend, frontend and mobile
> security and code review
> devops and infra
> data and AI
> even product and business roles
each one is a single md file. just copy it into .claude/agents and Claude can start handing it work!
pick the 2 or 3 that match your stack only tho, not all 166 or they start overlapping
way too many vibecoded apps go live without a single security scan. these 5 are free:
SECRETS
1. gitleaks/gitleaks: scans your repo AND its git history for keys (31k stars)
2. trufflesecurity/trufflehog: finds leaked keys and checks if they still work (29k stars)
CODE
3. semgrep/semgrep: catches injection bugs and risky patterns (17k stars)
PACKAGES
4. google/osv-scanner: flags dependencies with known holes (11k stars)
5. aquasecurity/trivy: packages, containers and configs in one scan (39k stars)
then paste this into Claude code:
"install gitleaks and osv-scanner, run both on this repo, explain every finding in plain english and fix only the real ones"
1
17
did you know you can point Claude Code at a model running on your own laptop now?
> ollama launch claude
thats the whole setup
→ local models cost nothing per token
→ cloud models are there if your machine is too slow
→ file edits, subagents and web search still work
it wont build what opus builds but for small fixes and quick scripts its plenty! :)
1
3
236
running Opus for everything is why a lot of people hit their claude limit by lunch
there's a model setting built for exactly this:
/model opusplan
→ in plan mode Opus does the thinking
→ once you approve, sonnet writes the code
→ the expensive model only works where it matters
2
2
430
that "yes, dont ask again" button in Claude code might have put your passwords on Github!
Apiiro found 2539 secrets across 1994 public repos, most in .claude/settings.local.json (the file where approved commands get saved)
30 sec fix:
1. add .claude/settings.local.json to your repo's own .gitignore
2. run: git log --all -- .claude/settings.local.json
3. if anything shows up, rotate every key in that file. deleting it now doesnt remove it from your history
19
why do people still pay for AI courses when Anthropic gives its own away on Github..?
> anthropics/prompt-eng-interactive-tutorial (39k stars)
step by step prompting course with exercises you actually run!
> anthropics/courses (23k stars)
API basics, real world prompting, prompt evals and tool use
> anthropics/claude-cookbooks (54k stars)
copy paste recipes for RAG, tool use, agents and more
> anthropics/claude-quickstarts (20k stars)
starter apps you can deploy from a support agent to computer use etc!
14
Anthropic, Vercel and many others give away the EXACT skills their agents run on FREE on GitHub:
1. anthropics/skills (192k stars)
Word, Excel, PDF and slides, frontend design, an MCP builder, even a skill that writes new skills
2. obra/superpowers (316k stars)
makes your agent plan first, write tests first, then build
3. vercel-labs/agent-skills (34k stars)
Vercel's own official collection
4. wshobson/agents (42k stars)
plugins, agents and skills written once and adapted for Claude Code, Codex, Cursor and Copilot
5. hesreallyhim/awesome-claude-code (58 stars)
the hand picked list of the best skills, agents and plugins people have shared
just install even one and see your output get helluva lot better!:)
links below👇
1
21