Ex-Quant trader at Citadel. Building the best book experience with AI
Joined July 2024
- Tweets77
- Following46
- Followers11
- Likes5
Grok Bot is sooooo smooth.... just get things done baby!
Claude is brilliant but slow, and it wants my attention ALL THE TIME. Config, skills, permissions, another file to tune before the work even starts. 30 minutes passed and it is still thinking about making the spreadsheet.
Grok Bot is fast and intuitive. Safer defaults than OpenClaw. Never gonna touch open claw again.
The gap isn't intelligence.
It's how humane it feels.
Wonderful job @elonmusk ! ALL in SpaceX at $111 :)
🤖 Made with AI
1
3
107
You can feel when a website was built by AI. Something's just… off. It blurs into every other AI site you've scrolled past.Same AI, same prompt, but the one with a proven design system made all the difference. Here's the exact file I wrote — and why every section matters 👇
1
2
88
Yesterday I switched from Claude Code to Codex mid-project, again due to Claude server issue. This becomes a normal state.
The handoff was smooth because the context wasn’t trapped inside one agent.
It lived in the system.
Here’s the portable context layer I’m trying to build:
1. Put agent instructions in the repo
AGENTS.md = shared operating manual
CLAUDE.md = Claude-specific adapter
README.md = what the project is / how to run it
2. Turn decisions into durable memory
Don’t leave product/technical decisions buried in chat.
Keep:
docs/decisions.md
docs/data-definitions.md
docs/current-state.md
3. Make work reproducible
Any agent should know how to:
run the app
refresh data
run tests
deploy
generate reports
If it takes a hidden manual step, write a script.
4. Own your data layer
User interviews, SQL queries, dashboards, feedback, outreach lists, analytics exports.
Put them somewhere inspectable.
The agent should retrieve context from your system, not hallucinate from vibes.
5. Add evals before swapping models
Can another model produce the same quality?
Does it cite the right data?
Does it preserve privacy?
Does it avoid false claims?
Model swapping without evals is just guessing.
6. Separate tools from truth
Tools, skills, and MCP servers are useful, but they are not the source of truth.
The source of truth should be:
repo docs
schemas
scripts
tests
dashboards
decision logs
7. Write handoff notes
At the end of a work session:
what changed
what works
what failed
what’s next
what needs a human decision
That one habit makes agent switching 10x easier.
The point is not to make Claude remember everything.
Or Codex remember everything.
The point is to build a system where context survives the agent.
Agents change.
Context compounds.
🤖 Made with AI
2
4
279
Claude becomes a super-power omni-function coworker! can't wait to try it.
Introducing Claude Tag, a new way for teams to work with Claude.
In Slack, Claude joins as a team member with access to the channels and tools you choose. Tag Claude in and delegate tasks to it while you focus on other work.
94
Hire a person or deploy an agent?
I’ve been thinking about this a lot. I think the answer depends on three things:
1. Know yourself: How do you want to spend your time? Managing people, or building systems?
2. Know your system
The bottleneck is rarely the person or the agent. It’s the quality of your knowledge, workflows, decision making, and feedback loops.
3. Know your goal
Why are you building in the first place? To make money? To create impact? For whom? What does success actually look like?
It is rarely about cost. Hiring the right person and building the right agent harness can be equally challenging.
2
52
Your intern now ships what your designer used to gatekeep.
Prompt: "create Instagram Carousel about how the book of Art of War shed wisdom for modern decision makers"
→ 90 seconds later: 4 on-brand slides, ready to post.
No Figma. No Canva. No "make logo bigger" Slack wars. No 3-day design-queue hell.
Brand-grade quality at typing speed.
AI doesn't make bad assets — bad systems do. Here's how we fixed it: 👇
1
1
112
Lazy brand systems + AI = garbage output.
Your brand guide is a 40-page PDF collecting dust.
Make it executable code.
We force the AI:
Hex codes baked into EVERY prompt
Fonts locked in the render engine
Voice as the system prompt
Logo rules as literal code (placement, sizing, no "guidelines")
AI stops guessing. Every output = on-brand by default. No review roulette.
1
30
🚨 NVIDIA dropped "AI is a Five-Layer Cake" at GTC 2026: Energy → Chips → Infrastructure → Models → Apps. But they skipped Layer 0: Raw Materials. No minerals = no AI. Countries are literally at war over this. China owns the mine. Wake up. #AI #Geopolitics #TechWar
1
1
197
2/7 China's iron grip (USGS 2025 / CSIS data):
Gallium: 98% global production (power electronics, fiber optics, defense)
Germanium: ~70% (semis, IR tech)
Rare Earth Processing: 90%
Neodymium magnets: 92% (motors, weapons, data centers)
Dec 2024 export ban to US on gallium/germanium? Estimated $3.4B GDP hit. This isn't theory—it's happening.
1
1
20
I ignored my family on Chinese New Year eve to set up @openclaw on a dedicated Mac Mini.
Husband: “You’re doing WHAT tonight?”
Me: “Just 3 hours… trust me.”
3 weeks later: this thing quietly became my calmest, most powerful co-founder.
Honest review 🧵 (setup warts + real magic) 🦞 openclaw.ai
1
83
Setup reality check (not the shiny docs version):
curl install worked → then silence. No obvious “add your API key here” step.
Had to hunt: openclaw onboard --anthropic-api-key $KEY
Dashboard? Accidentally found http://127.0.0.1:18789/
Whatsapp and Telegram took 1 hour of trial & error.
@openclaw team — first-time UX needs love. Still worth it.
Pro tip from someone who learned the hard way: → Dedicated Gmail account just for Claw. Cleaner, safer. → Use a dedicated cheap Mac Mini if you can. Feels like giving your AI its own office.
1
5
Every major leap in human history follows the same pattern:
Exponential growth → constraint → phase change → new rules.
AI is not an exception.
It’s just faster.
1
1
62