side project: using claude to scrape and generate for me past historical trade performance

trying to build this further as a pilot to filter out quality of alpha trade calls on certain individuals I am tracking.

It made me miss 2024, when it was so much easier to trend trade 😭

Can we bring biden back?

Unsloth Studio is a free, open-source tool that finally lets normal people create and customize their own AI chatbot right on their laptop.. no coding needed.

• No-code fine-tuning (upload your PDFs/CSVs/docs → auto-generate training data → train LoRA/fine-tune models 2× faster + using ~70% less memory)

• Pick a free open model (like a smarter, private version of ChatGPT) → teach it your style, your work knowledge, or your hobbies in a few clicks
Chat with it, compare versions side-by-side, and save your custom AI forever — all offline, private, on your Mac/Windows/Linux computer (even older ones work surprisingly well!)

Think of it like: instead of renting someone else's AI, you now build and own your personal version that really knows YOU. No cloud, no data leaks, no subscriptions.If you've ever thought "I wish this AI understood my notes/emails/job better"… this tool makes that possible without being a programmer.

https://x.com/UnslothAI/status/2033926272481718523

Lel this is so cute

Created an overview of Claude Code: Kawaii Edition

hope you like it

🛠 Superpowers: Agent Skills Framework

This GitHub has been trending recently

Tired of AI coding assistants that jump straight into code and miss the point?

Superpowers is a set of "skills" - instruction templates that make agents plan before they build.

What it changes:

❌ Before: "Build me a bot" → 500 lines of code → you review and find bugs

✅ After: "Build me a bot" → agent asks questions → shows design → you approve → builds in small tasks with tests

The skills:

• Brainstorming — refine ideas through questions
• Planning — break work into 2-5 min tasks
• TDD — tests first, always
• Subagent dev — fresh agent per task, self-reviews
• Git worktrees — clean branches, easy rollback

Works with: Claude Code, Cursor, Codex, OpenCode, Gemini CLI

Why it matters: Agents stop doing their “own thing” They follow process. You approve plans, not just code.

🔗 github.com/obra/superpowers

Worked further on my live alpha feed from various sources direct to Telegram.

Added an additional feature to AI summarise images/videos so I save time and get the TLDR(s).
Used Gemini flash for the transcripts (pretty good!)

problems I faced:

• getting the formatting rules for telegram right and reverse-engineering the feeds.

• sync locks with my telegram / database

• hitting telegram bot/rate limits

• claude code losing context (used the app, and the session just disappeared magically after restarting the app, should have just stuck with CLI)


This was a real pain in the ass to do, had to rebuild this >4 times, while waiting for claude limits to cooldown

A breakdown of Claude simplified for the non-techie people <33

https://x.com/vincent_koc/status/2035824467805606010

Incoming 👀🦞

one week left to tokenmaxx (2x ur usage) ur claude subscription while it lasts.

take note of the hours.

(8pm - 2am) SGT

Openclaw Big update 🦞🦞

breaking changes:

openclaw 2026.3.22 changes:

• clawhub marketplace — app store for plugins and skills
• per-agent reasoning — each agent can have its own thinking level
• /btw command — ask quick questions without losing your place
• ssh sandboxes — run agents on remote computers
• more search options — exa, tavily, firecrawl added
• better chinese ai — minimax m2.7
• cheaper openai options — gpt-5.4-mini and nano
• chrome extension gone — use browser directly instead
• image generation changed — new way to set up image models
• matrix chat improved — better matrix support

full changelog: github.com/openclaw/openclaw/releases/tag/v2026.3.22

Warning: Dont update to the latest openclaw. it bricked my setup. wait for a more stable release

Saw a non-techie mom code this for her kids to practice chinese character strokes and memory (lel i forgot there was even an order for strokes)

Super cool, thought it’s worth sharing

https://claude.ai/public/artifacts/c9ad0b3f-1d26-4642-8c49-d9d111bd3803

Wanted to celebrate this small little win with a friend/community member

ZERO technical background, had no idea what claude was but decided to join the war room to do the claude partner sprint together as a grp, and used his new skills to land himself a sales lead

Crazy stuff

Claude just shipped Claude Computer

It can now practically just do whatever you are doing with mouse functions as well

• open apps on your mac
• navigate browsers
• fill in spreadsheets
• click buttons, type text
• move files around
• basically anything you'd do manually at your desk

https://x.com/claudeai/status/2036195789601374705

Wow.. OpenAI sunsetting their video arm.
Maybe all that compute for AI slop wasn’t worth it. Consolidation into their codex/chatgpt/browser

Google’s probably the incumbent to inherit their user share?

https://x.com/soraofficialapp/status/2036546752535470382?s=46

Insane. everything Claude shipped in the last 52 days.

Have been skimming some through some AI video/image gen mass output recently, learning about AI influencers, etc. and went philosophical again.

spitballin some thoughts here:

There’s so much AI slop/fake engagement, these days, I wonder how long before the mirage fails. I get DMs daily from friends asking me if this is real or not.

Let’s just put this this way:

Your Input = Your Output
Domain Expert Input = Domain Expert Output

Even if you reverse engineer things - you are still trying to reverse engineer with your own train of thoughts.

If you don’t have the experience of a subject domain expert (like a videographer or brand agency owner with 10 years of actual design experience), there are ceiling limits to the prompt outputs. The best way to cover that gap isn't more prompting, it's understanding the domain expertise before trying to prompt the LLM with limited knowledge.

Insanity is doing the same thing over and over, expecting different results said some mad scientist.

The LLM is just a glorified token vending machine, giving you what you want based on the input.

They are probabilistic, not "thinking" in the human sense. They predict the next token based on the patterns in your prompt. If the input lacks depth, the probability path the AI takes will lead to a generic, "mid-tier" output.

It’s not about coding plugins or trying to disrupt a business with AI - it’s the expertise gap. Business owners know AI exists, but can they trust you? Anyone can generate AI.

The reputation layer: ur track record, delivery history, accountability, is what’s missing.

We are moving into an era where Digital is for work, and Analog is for life.

The more "perfect" and algorithmic the world becomes, the more the wealthy will pay for the "beautifully broken", the wabi-sabi and for artisanal craft.

If there are "50 wrappers" for a tool like OpenClaw, the technology itself is no longer a moat. It’s becoming a commodity, like electricity. When everyone is "super accelerated," the speed of the tool stops being a competitive advantage. The digital is now the floor, not the ceiling.

Value always flows toward what is scarce.

Two takeaways in parallel: use digital for leverage, analog for identity

Digital for Leverage: Use the "industrialized" AI to handle the high-volume, repetitive work. This is the engine that buys back your time.

Analog for Identity: Invest that reclaimed time into things that "pass the test of time." This is the real life - ur family, enjoyment of hobbies, working out, being a blessing to others

Very soon you will be able to use your own LLMs for Siri. Good move Apple.

https://x.com/markgurman/status/2037230804942610548?s=46