Vibe Science ✨
Claude · Codex · Gemini · LabClaw

# Install via npm (Node.js required)
npm install -g @anthropic-ai/claude-code

# Or via pip wrapper
pip install anthropic

# Set your API key (get one at console.anthropic.com)
export ANTHROPIC_API_KEY="sk-ant-..."

# Verify
claude --version

# Install Codex CLI
npm install -g @openai/codex

# Or use the OpenAI Python SDK directly
pip install openai

# Set API key (platform.openai.com)
export OPENAI_API_KEY="sk-..."

# Verify
codex --version

# Install Gemini CLI (Google AI SDK)
npm install -g @google/gemini-cli

# Or Python
pip install google-generativeai

# Set API key (aistudio.google.com — has a FREE tier!)
export GEMINI_API_KEY="AIza..."

# Verify
gemini --version

# Gemini 1.5 Pro has a free tier with 1M context window
# You can literally paste a GTF file and ask questions about it

# Ask Claude a direct question
claude "What does a TSS enrichment score of 1.2 mean for my ATAC-seq QC?"

# Pipe a file to Claude for analysis
cat alignment_stats.txt | claude "Summarise these alignment stats and flag anything concerning"

# Ask Claude to write a script
claude "Write a Python script that reads a VCF file with cyvcf2,
filters to PASS variants with MAF > 0.01, and outputs a tidy TSV
with columns: chrom, pos, ref, alt, gene, consequence, AF"

# Have Claude review your existing script
claude "Review this DESeq2 script for statistical mistakes" < deseq2_analysis.R

# It's 2am. STAR alignment crashed. You have no idea why.
# Old you: Stack Overflow for 90 minutes
# New you:

cat star_log.final.out | claude "Why did my STAR alignment fail and how do I fix it?
I am running human GRCh38 with splice junctions. Here is the log:"

# Claude will calmly identify that your --genomeSAindexNbases was wrong for your genome size
# and give you the exact corrected command.
# You will feel both grateful and slightly embarrassed.

# Start an interactive session
claude

# Then inside the session:
# > I have 48 RNA-seq samples, 3 conditions (control, drug_A, drug_B),
# > 16 replicates each, human samples. I want to do differential expression
# > analysis. Walk me through the best approach step by step.

# Claude will ask clarifying questions, recommend DESeq2 vs edgeR based on
# your design, warn you about batch effects, and generate the full R script.
# This used to take a week to figure out from papers. Now: 8 minutes.

from openai import OpenAI

client = OpenAI()

# Use the API programmatically in your analysis notebooks
def ask_codex(prompt, system="You are an expert bioinformatics engineer."):
    response = client.chat.completions.create(
        model="gpt-4o",
        messages=[
            {"role": "system", "content": system},
            {"role": "user", "content": prompt}
        ]
    )
    return response.choices[0].message.content

# Example: generate a parsing function on the fly
code = ask_codex("""
Write a Python function parse_blast_xml(filepath) that:
1. Parses BLAST XML output using BioPython
2. Returns a pandas DataFrame with columns:
   query_id, subject_id, pident, length, evalue, bitscore, query_start,
   query_end, subject_start, subject_end
3. Filters to evalue < 1e-5 and pident > 70
""")

print(code)
# Paste or exec() the output — instant working parser

# All those tasks you've been putting off for months:

codex "Write an argparse CLI wrapper for my RNA-seq pipeline with
flags: --input-dir, --output-dir, --genome, --threads, --dry-run,
--config, --resume. Include --help text for each flag."

codex "Convert my hardcoded paths in analysis.py to a config.yaml
system using pydantic for validation"

codex "Write unit tests for my VCF parsing functions using pytest
with 3 synthetic VCF fixtures"

# Every bioinformatician has a list of 20 scripts they need to write.
# Codex writes all 20. You drink your coffee.

import google.generativeai as genai
import pandas as pd

genai.configure(api_key="YOUR_GEMINI_API_KEY")
model = genai.GenerativeModel("gemini-1.5-pro")

# Load your DESeq2 results
df = pd.read_csv("deseq2_results.csv")

# Automated biological interpretation
prompt = f"""
I have DESeq2 differential expression results comparing cancer vs normal tissue.
Here are the top 50 significant genes (padj < 0.05, |log2FC| > 1):

{df[df.padj < 0.05].nlargest(50, 'log2FoldChange')[['gene_name','log2FoldChange','padj']].to_string()}

Please:
1. Identify major biological pathways involved
2. Highlight any known oncogenes or tumour suppressors
3. Suggest the most relevant Gene Ontology terms to test for enrichment
4. Flag any surprising findings
"""

response = model.generate_content(prompt)
print(response.text)

# This replaces about 45 minutes of manual literature lookup
# Your PI asks "so what do these genes do?"
# You already know because Gemini told you 3 minutes ago

# You send ONE message to your OpenClaw agent:
→ USER: "Run a complete scRNA-seq analysis on the 10x data in /data/pbmc_10x/.
         Cluster the cells, annotate them, and find differentially expressed genes
         between CD4 T cells and CD8 T cells. Save all plots to /results/."

# The agent:
# 1. Loads the LabClaw tooluniverse-single-cell skill
# 2. Runs Cell Ranger count if raw FASTQs are present
# 3. Imports into AnnData / Seurat automatically
# 4. Applies QC filters (adaptive thresholds)
# 5. Normalises, scales, PCA → UMAP
# 6. Clusters with Leiden (tests multiple resolutions)
# 7. Annotates cell types using SingleR / CellTypist
# 8. Runs FindMarkers for CD4 vs CD8
# 9. Generates volcano plots, UMAP PDFs, marker heatmaps
# 10. Writes a structured summary report
# All without another message from you.

# You get back:
← AGENT: "Analysis complete. Found 14 cell types across 8,423 cells.
          Top DE genes CD4 vs CD8: CD4, IL7R, CCR7 (up in CD4);
          CD8A, GZMB, PRF1 (up in CD8). Results saved to /results/.
          Report: /results/analysis_summary.html"

# Each LabClaw skill is a structured YAML/Python definition:
# Format: Overview → When to Use → Capabilities → Examples

skill_name: tooluniverse-single-cell-rna
domain: biology
when_to_use:
  - "User has scRNA-seq data (10x, Smart-seq2, Parse, etc.)"
  - "User wants QC, clustering, annotation, or DE analysis"
  - "Any mention of Seurat, Scanpy, Cell Ranger, or AnnData"
capabilities:
  - Load CellRanger output or h5ad/RDS files
  - QC filtering with adaptive thresholds
  - Normalisation (SCTransform, scran, log-normalise)
  - Dimensionality reduction: PCA, UMAP, tSNE
  - Clustering: Leiden, Louvain
  - Cell type annotation: SingleR, CellTypist, scGPT
  - Differential expression: FindMarkers, DESeq2 pseudobulk
  - Visualisation: UMAP, violin, dot, heatmap, volcano
examples:
  - "Analyse PBMC 10x data and find all immune cell types"
  - "Compare CD4 vs CD8 T cells in tumour microenvironment"