Skills pack · Data quality

The data quality skills pack

Point Claude Code at your own contact export and have it counted properly: duplicates, malformed emails, empty columns, inconsistent formatting, stale records. Reads CSV and multi-sheet Excel. Runs on your machine, nothing uploaded.

The checklist tells you what to look for. This does the counting.

It is an agent skill: unzip it into .claude/skills/, start a session, and point it at a file. It audits the export, explains what it found, and helps you decide what to do about it. There is nothing to sign up for and nothing to install.

Why it runs on your machine

Handing a contact database to a company you are still evaluating is a bad trade, and you should not have to make it to find out whether your data is in trouble. So the audit comes to the file rather than the other way round. Your export is opened read-only, nothing is uploaded, and nothing is written unless you ask for a flagged copy. You can read every line of the script before you run it.

What it checks

Duplicates, using the rules GuardenerAI seeds into a new workspace, ported line for line. Exact-normalised email, which the product treats as near-certain and blocks on. First name plus last name plus digits-only phone, which it treats as a strong signal and flags for a human. Company names with Ltd, Inc and the rest stripped before comparison, so Acme Dental and Acme Dental Ltd stop being two accounts.

Email syntax, including the role addresses standing in for a named person. It does not claim to know whether an address still receives mail, because a file cannot tell you that.

Fill rate per column, so you can see which fields are actually empty instead of being told that “required fields are missing”. Which ones matter is your call.

Staleness, but only when your export has a date column it can read. When there is not one, the report says so rather than scoring a dimension it could not measure.

Formatting, meaning the phone shapes and the company spellings that make a second de-duplication pass find records the first one missed.

What it will not do

It will not verify deliverability, guess a missing value, enrich from an outside source, or decide on its own that a record is dead. Those need a person or a system with more context than a spreadsheet has. It also will not merge anything without showing you the cluster and asking which record survives.

The honest limitation

This is a point-in-time audit of a file you exported. It finds what is already wrong. It cannot stop the next duplicate from being created five minutes after you finish, because by then the data is back in whatever system you exported it from. That part is the difference between an audit and a CRM that applies the same rules at the point of entry, which is the only moment a duplicate is cheap to stop.

Running it on your own export is a reasonable way to find out whether that difference is worth anything to you.

Keep going

Related reading