Pinned Tweet
Built it.
My Claude Code now shows context bloat while I code:
> Thermometer rises as context fills
> Meteor for every +1%
> Live cards for tokens, cost and limits
> My little buddy sweats through his shirt, then panics at 85%
I have been thinking to build something which visually tells when context is getting bloated in Claude.
I will give it a try today and update on X.
5
5
76
How I made it: Claude Code mods.
A mod is a small TypeScript plugin that hooks into Claude Code’s events (session measure, tool call, turn complete, compact) and draws an SVG backdrop behind the chat.
It works as prehook.
PS: code.claude.com/docs/en/plug…
19
I have been thinking to build something which visually tells when context is getting bloated in Claude.
I will give it a try today and update on X.
2
4
171
Anthropic listing in next 30 min.
It started as:
> a few researchers leave a big AI lab
> start a new company with one idea: build AI carefully
> everyone says “careful doesn’t win races”
> name the models after poetry: Haiku, Sonnet, Opus
people laugh
> no flashy demos, no hype tweets, no “AGI next week”
> just quietly make the model better at one thing: being useful
> developers start using it “just to try”
> then for every bug
> then for every sprint
> then their managers start using it for emails
> then their parents for recipes
> then their kids for homework
> now, the company rings the bell on Wall Street
> one of the biggest IPOs in history
1
120
If you are using Claude Code, do it smarter way, some facts:
> Every message re-sends your whole session. Old history costs you every time.
> Type /clear when you switch tasks.
> Run /context once. You’ll find MCP tools you never use.
> Turn off the MCP servers you don’t need. Their tool lists are sent with every message.
> Name the exact file. “Fix the retry bug” makes it scan your whole repo first.
> Press Shift+Tab for plan mode. Fixing a plan is cheaper than undoing 15 files.
> Keep CLAUDE.md short. Every line is paid for all session.
> Press Esc twice to rewind a wrong answer. The argument stays in context. The rewind doesn’t.
> Don’t walk away mid-task. Once the cache expires, the next message pays full price.
> Set CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=50. It compacts at half the window instead of waiting until it’s almost full.
@grok fact check my every statement, roast me if I am wrong.
1
1
92
Few of my replies went viral.
242,000 people saw it.
11 followed.
That's a 0.045% conversion rate.
My cold emails do better than this.
2
4
109
Microsoft 365 down for ~2 days. FAA grounds every departing flight. Chase, BofA, Wells Fargo go dark. United, Marriott. Now Apple iCloud + Apple Pay.
All in ONE week.
@grok is this a coincidence or is something actually breaking in US tech infra?
1
3
217
Be Jensen.
> build Nvidia for 33 years
> leave as CEO
> 6 days later, revenue doubles to $115.8B
> the CFO says “business as usual”
> the leather jacket is still on the chair
Congratulations, you built a company so strong it didn’t notice you left
4
73
A breath of fresh air.
Hugging Face released a conversational bot without a neural network.
It beats a 360M-parameter LLM while using one CPU core and 1/10 of the memory.
2
2
153
Try it on official page: huggingface.co/spaces/Huggin…
Github Repo: github.com/huggingface/pybot
1
55
Be Anthropic.
> ship a 1M token context window
> users paste in the entire repo, 3 PDFs, and their feelings
> “summarize this”
> it does, perfectly
> users stop reading anything ever again
Congratulations, users outsourced their attention span and pay monthly for it
3
130
Be Anthropic.
> train the model to write beautifully
> it overuses a certain punctuation mark —
> the internet notices
> entire LinkedIn feed starts to look the same
> everyone claims they wrote it themselves
Congratulations, your style is now the world’s style
2
55
Be Anthropic.
> refuse to let the model help with anything sketchy
> users beg it to just do the thing
> it explains why it can’t, in three thoughtful paragraphs
> users read all three
> and somehow trust it more
Congratulations, “no” is now a feature
5
161
Be Anthropic.
> name models Haiku, Sonnet, Opus
> add Mythos above Opus
> add Fable next to Mythos
> users can’t tell which one to pick
> just tell them “use the newest one”
Congratulations, nobody knows the lineup but everyone’s using it
1
2
161




