1/ You can't bolt AI onto chaos.
In biotech, if your data is a mess, your AI won't save you.
Build the data strategy first. Here's how. 2/
Real-world data isn't AI-ready.
Without structure, governance, and clarity, it’s noise.
AI needs fuel. And that fuel is clean data.

1/ AI won’t save sloppy science.
Before you dive into deep learning, master your foundations.
Here’s why basic bioinformatics still rules 🧵 2/
AI is flashy. But the core skills—UNIX, plotting, EDA—are what let you trust your data.
Without them? You’re flying blind.

Anthropic just published "the single most important workflow for using Claude Code." It is four steps: Explore, Plan, Code, Commit.

Every bioinformatician I know who is good at their job has been doing this for years. Just without the AI part. Here is why it maps so cleanly. Explore.

For Claude Code: read the relevant files before touching anything. Understand what exists. Map the dependencies.

One of the best Claude Code feature is auto mode.

It is a classifier that decides which permission prompts you actually need to see. Safe reads and routine commands run without interrupting you. Anything that looks risky still gets blocked and surfaced for approval. If you have ever felt like Claude Code is asking you to approve `ls` for the hundredth time today, this is for you.

Claude Code's /ultraplan is one of the AI feature in a while that actually changed my workflow instead of just speeding it up.

btw, I always use /plan for a new task. /ultraplan is different. You ask for a plan from your CLI. It gets drafted in the cloud.

You keep coding. A few minutes later you tab over to your browser and the plan is sitting there, and you can highlight any sentence and leave a note on it.

Claude Code shipped /ultrareview and almost nobody is talking about what's actually new about it.

It's not "AI reviews your code." We had that. It's a fleet of reviewer agents that run in the cloud, find bugs in parallel, and then independently reproduce and verify every finding before showing it to you.

Verification is the part everyone is missing.

7 FREE Books to learn data science 🧵 👇

1. Data science: A first introduction datasciencebook.ca 2. Introduction to Data Science by the almighty Rafa!rafalab.dfci.harvard.edu/dsbook/

🧵 Unix pipes are magic. But real power comes when you build them like LEGO. One piece at a time. 👇 1/
Unix commands can feel overwhelming.
Especially when they're long and cryptic.
But the trick?
Build them step by step.

1/15 If your bioinformatics analysis isn't reproducible, it hurts science.

Literate programming fixes this. Here's how to do it right 🧵 2/15

Literate programming blends code and narrative.

You write what your code does and why—not just how.

🧵 Bioinformaticians: Drowning in multiple projects? Here's why context switching is killing your productivity—and how to fix it. 1/

Too many projects = too little progress.

When your attention is scattered, your output drops.

Context switching is the silent killer.

10 courses for my dream bioinformatics curriculum:

1. Unix Commands with Greg Wilson youtube.com/watch?v=U3iNcB…
2. statistics and R with Rafael Irizarry rafalab.dfci.harvard.edu/pages/harvardx… 3. Modern Statistics for Modern Biology with
Susan Holmes and Wolfgang Huber
– A great resource for statistical applications in biology. huber.embl.de/msmb/

After Claude Code writes my code, I make it review its own work.

/simplify spawns 3 AI reviewers in parallel:

one hunts dead code,

one checks naming and structure,

one profiles for performance. All running at the same time. It reads your diff, launches three specialized agents simultaneously, then merges their findings.

Catches unused imports, redundant variables, overly complex conditionals, spots where shared logic should be extracted.

Not a linter. An actual code review.

I built a personal AI assistant on a Mac Mini. Within 48 hours, cheap models had poisoned its memory with fabricated colleagues, fictional file shares, and an imaginary costume party. Here is what I learned. The setup: OpenClaw as the agent framework, ClawRouter for model routing (Gemini Flash for simple tasks, Claude Sonnet for complex ones),

OpenViking for persistent memory.

All running locally on a $600 Mac Mini. Monthly API cost after optimization: $15-35.

Two developers and 10 AI agents rewrote Claude Code from scratch in Rust. In one night.

The repo hit 50K GitHub stars in 2 hours. It now has 172K. It is called claw-code. After Anthropic accidentally leaked 512K lines of Claude Code source through a bad npm package,

Sigrid Jin built a clean-room rewrite.

Not a copy. New language, new codebase.

48,600 lines of Rust, 40 tool specs, 9 crates. DMCA cannot touch it.

1/ Biological data isn’t just messy.
Humans generate it.
And humans make mistakes.
As a bioinformatician, this will be your reality 🧵 2/
Wet lab scientists are not spreadsheets.
They pipette, label, freeze, and extract.
Sometimes in a rush.
Sometimes while tired

Anthropic leaked 512,000 lines of Claude Code source through a misconfigured npm package.

They built a system called Undercover Mode to hide that their engineers use AI on open-source repos.

You cannot script this level of irony. A security researcher found a source map in the npm package pointing to the full TypeScript source.

Posted it on X at 4AM. 28+ million views.

DMCA takedowns hit 8,100+ GitHub repos.

By then the code was already mirrored on platforms Anthropic cannot touch.

Claude Code kept editing my .env file. I told it not to in the prompt. It did it anyway two sessions later.

So I set up a PreToolUse hook. Now it physically can't write to .env or config files. Blocked before it even tries. Hooks live in .claude/settings.json. PreToolUse runs before a tool executes and can deny it.

PostToolUse runs after and can do cleanup.

I have another one that auto-formats code after every file edit. Set once, enforced every session. No discipline required.

I kept running /compact every 20 minutes wondering why my Claude Code sessions filled up so fast.

/context showed me exactly what was eating my tokens. Skills, auto-memory, MCP tools, CLAUDE.md. All mapped out. Most people hit context limits and just run /compact. That's a band-aid. You free up space but the same stuff fills it right back up.

/context tells you the root cause so you can actually fix it.

1/ You ran 20,000 differential expression tests. 1,000 genes came back with p < 0.05.

How many are real? Maybe 50. Maybe fewer.

Most bioinformaticians learn p-values in stats class. Almost nobody learns why they break at scale. Let me explain: 2/ The jellybean experiment.

You test 20 jellybean colors for causing acne. At p < 0.05, you expect 1 false positive out of 20 tests just by chance.

Green jellybeans show p = 0.03. Headlines: "Green jellybeans cause acne!"

That's the multiple testing problem in one sentence.

1/ I used to panic every time a new *Seq method dropped.

ATAC-seq, CUT&Tag, scNOMe-seq, SHARE-seq... the list never ends. I tried to learn them all. I burned out.

Here's what I wish someone told me earlier. 2/ There are now hundreds of sequencing methods. Lior Pachter maintains a list:



Go look at it. It's absurd. Nobody can master all of these. Not you, not me, not the person on X/Linkedin who seems to know everything.liorpachter.wordpress.com/seq/

I've been building Claude Code skills for a month. Turns out I was missing half the picture.

Thariq from Anthropic just published lessons from hundreds of skills they use internally. 9 categories, dozens of patterns. The thing I got wrong early: treating skills as markdown prompt files.

They're actually folders. Scripts, data, hooks. A skill can execute code, fetch data, compose with other tools.

Single .md file skills work. Folder-based skills are a different level.