I've spent the past year integrating LLMs into my daily bioinformatics workflow — mostly RNA-seq, single-cell, and spatial transcriptomics. Here's what actually stuck:

1. LLMs are best at the boring parts.

Writing boilerplate Seurat/Scanpy code, reformatting annotations, generating QC reports. The time savings compound fast. Where they fall flat: interpreting biological significance. You still need domain knowledge.

2. Reproducibility gets harder, not easier.

When your analysis pipeline includes natural language prompts, version control becomes a real problem. I now track every prompt alongside my code — same commit, same repo. It's the only way to keep things auditable.

3. The real value is in structured exploration.

LLMs are surprisingly good at suggesting analysis directions you hadn't considered — alternate clustering approaches, unusual gene set combinations, spatial neighborhood definitions. The key is treating them as a brainstorming partner, not an oracle.

I'm writing up my full workflow in a weekly article series here. The first one (on setting up a reproducible LLM-integrated computing environment) is free if you want to see the approach.