How to use the RDKit Cheminformatics Guide agent skill

What does the chemistry-rdkit skill make the agent do?

An agent skill is a folder containing instructions that an AI agent reads when a task matches the skill's description. This one contains a single file, SKILL.md, with reference guidance for RDKit. It is guidance text, not a standalone tool: the agent still writes and runs its own Python code.

When a request involves SMILES strings, drug-like compounds or structure comparison, the agent has a reference for loading molecules, checking for invalid input, calculating properties such as molecular weight and LogP, comparing molecules by fingerprint similarity and searching for functional groups.

• Load molecules from SMILES, SDF and SMILES files, and write SDF output
• Calculate descriptors such as molecular weight, LogP, TPSA, hydrogen-bond donors and acceptors, and rotatable bonds
• Apply the Lipinski Rule of 5 thresholds (MW <= 500, LogP <= 5, HBD <= 5, HBA <= 10)
• Compare molecules with Morgan, RDKit and MACCS fingerprints using Tanimoto similarity
• Find functional groups with SMARTS patterns

How do you install the chemistry-rdkit skill?

You can install the skill from the ZIP file on this page in one of two ways. Use the folder option if you work with Claude Code, or the upload option if you use the Claude apps.

For a personal install, unzip the download into ~/.claude/skills/. For a single project, unzip it into the project's .claude/skills/ folder. In the Claude apps, open Settings, go to Capabilities, and upload the skill under Skills.

• Personal: unzip into ~/.claude/skills/ so the skill is available everywhere
• Project only: unzip into .claude/skills/ inside your project folder
• Claude apps: Settings → Capabilities → Skills, then upload the skill

Install RDKit so the examples can run

The skill does not install anything itself. It needs Python and the RDKit library. You can install RDKit yourself, or ask the agent to install it as part of the task. The examples in the skill cannot run without both.

• Python must be available on your machine
• The RDKit library must be installed in that Python environment

How do you trigger the skill with natural requests?

You do not need to name the skill. The skill's description covers SMILES, drug discovery and cheminformatics tasks, and its trigger keywords include molecule, SMILES, chemical, drug, rdkit, fingerprint, molecular, compound, reaction and cheminformatics. Plain requests that use these words are the most likely to load it.

Be specific about the input and the outcome you want. Say which molecules you have and what you need back.

• "Calculate molecular weight, LogP and TPSA for these SMILES: CCO, c1ccccc1O"
• "Check which compounds in this SMILES list pass Lipinski's Rule of 5 and list any that fail."
• "Rank my library by Tanimoto similarity to this reference molecule using Morgan fingerprints."
• "Find all carboxylic acids in this SDF file using a SMARTS pattern."

What does a typical result look like?

For a property request, expect the agent to parse each SMILES, skip or report entries that return None, and print the descriptor values. For a similarity request, expect Morgan fingerprints at radius 2 with 2048 bits and Tanimoto scores for each compound. The skill notes that Tanimoto above 0.7 suggests structural similarity in a virtual-screening context.

Check the output yourself. The skill's rules of thumb are screening heuristics, not validated predictions.

What safety and practical limits should you know?

The skill's metadata says it does not run scripts and does not use the network. It requires Python and RDKit. Because it is only guidance, the code the agent writes is what runs on your machine, so review it before running it on data you care about.

The skill does not address medical, safety or clinical decisions, synthesis, hazardous chemicals, lab handling or regulation. Descriptors and filters such as Lipinski and Veber are screening aids, not predictions of toxicity or efficacy.

• Do not use the output for medical or clinical decisions
• Do not rely on it for hazardous-chemical or synthesis work
• Check that MolFromSmiles did not return None before trusting results

Tips for better results

Give the agent clean input. Invalid SMILES is the most common source of trouble, and the skill advises checking for None and treating kekulization errors as a sign of invalid input. Ask the agent to report rejected entries rather than silently dropping them.

Use the defaults unless you have a reason not to. The skill recommends sanitizing molecules by default and using radius 2 for Morgan fingerprints for most applications. For 3D work, add hydrogens first, and for large SDF files, use ForwardSDMolSupplier to save memory.

• Ask for the count of valid and invalid molecules in every batch
• State the fingerprint and threshold you want, or accept the defaults
• For large files, mention the file size so the agent can stream it