Hello Bioinformatics lovers,

The AI coding assistant toolkit is advancing fast.

You might have heard about github copilot or Claude Code.

I urge you to try those new things and learn what they can and can not do.

I built this daily quote app with Claude Code in 30 mins and deployed it on Vercel https://daily-quote-app-mu.vercel.app/​

I also built this Free interactive R and Python learning app during the holidays https://chatomics-labs.base44.app/​

Keep an open mind!

It was impossible for me years ago.

I am not a software engineer, but now the barrier is so much lower.

Yeah, I know you subscribed to this newsletter for learning bioinformatics, but it applies to learning anything in life :)

​

I am using Claude code as an example in this newsletter, but the skills apply to any AI-coding tools.

Most bioinformaticians blame Claude Code when their genomics analyses fail.

The real problem? You're treating it like Stack Overflow when you should be treating it like a junior analyst.

Here's what actually works.

Give context, not just commands

The difference between mediocre and exceptional results comes down to your first prompt.

What doesn't work: "Analyze this RNA-seq data"

What does: "I have 12 tumor samples vs 12 normal controls, paired-end 150bp Illumina sequencing. I need differential expression analysis with DESeq2. Samples are already batch-corrected for sequencing lane."

Claude needs the biology and the technical details. Both matter.

Upload everything, describe nothing

Stop trying to summarize your data. You're not saving time—you're losing accuracy.

Your GTF has non-standard formatting? Upload it. Your sample metadata has a complex structure? Upload it. MultiQC flagged quality issues? Upload the report.

Claude can read files faster and more accurately than you can describe them. Let it.

Plan before executing

The best analyses start with strategy, not code.

Try this: "I need to identify differentially accessible chromatin regions from ATAC-seq. Walk me through the analysis plan before writing any code."

Claude will map out:

• Peak calling strategy
• Normalization approach
• Statistical testing method
• Multiple testing correction
• Visualization plan

Review the plan. Then say "Run it."

This catches conceptual errors before they become computational ones.

Share your expectations

Your domain knowledge is Claude's missing piece.

Tell it what normal looks like:

"I expect 1,000-5,000 peaks near transcription start sites. If you find 50,000, something's wrong with the peak calling parameters."

"Batch effects are my biggest concern. Check for them before running differential expression."

"I need FDR < 0.05 AND log2 fold-change > 1 for biological significance."

These guardrails prevent biologically meaningless results that are technically correct.

Use checkpoints, not full pipelines

Complex analyses need waypoints.

Instead of: "Run the entire single-cell pipeline"

Try this:

1. "Load data and show QC metrics" → review
2. "Remove low-quality cells and doublets" → review
3. "Cluster and show marker genes" → review

Each checkpoint is a chance to catch errors early, when they're easy to fix.

Demand explanations

When something looks suspicious, ask why.

"Why did you use TMM normalization instead of DESeq2's default median-of-ratios?"

"Why filter genes with fewer than 10 counts?"

"Why use resolution 0.8 for Louvain clustering?"

If Claude can't justify its choices biologically, question the result.

Iterate with precision

When results miss the mark, give specific feedback.

Vague: "Do it better"

Specific: "These peaks are too broad. MACS2 is probably merging nearby signals. Re-run with the --call-summits flag and --extsize 200."

Precise feedback gets precise fixes.

The workflow that actually works

Here's the pattern I use for every analysis:

1. Upload all data files (no summaries)
2. Ask Claude for an analysis plan
3. Review the plan for biological sense
4. Execute in checkpoints
5. Review each checkpoint before proceeding
6. Ask Claude to explain anything suspicious
7. Iterate with specific, technical feedback

The key insight: You make the biological decisions. Claude executes them.

That division of labor is what separates mediocre results from publication-quality analysis.

What this changes

Once you stop treating AI coding tools as oracles and start treating them as capable but naive analysts, everything shifts.

You're no longer fighting the tool. You're directing it.

And your results show it.

What's your experience with AI coding tools for genomics? Hit reply and let me know what's worked (or hasn't) for you.

Happy Learning!

Tommy aka crazyhottommy

PS:

If you want to learn Bioinformatics, there are other ways that I can help:

1. My free YouTube Chatomics channel, make sure you subscribe to it.
2. I have many resources collected on my github here.
3. I have been writing blog posts for over 10 years https://divingintogeneticsandgenomics.com/​
4. Always read my Linkedin posts​​

Stay awesome!