Claude Skill for Database Exports: Make Sense of Hundreds of CSVs, No Code Required

A folder of CSVs and a bad feeling.

Every migration starts the same way. The client sends a zip file, you unzip it, and you are looking at 47 CSVs with names like tbl_CONTACT_ADDR_X.csv and no documentation. Somebody in 2011 knew what ParentRefID pointed to. That somebody is retired.

You could open each file in Excel, squint at the columns, guess the keys and build a data dictionary by hand. I did that for years. It takes a day or two per source system, and you still miss things, because a human cannot eyeball 200,000 rows to confirm that a column is actually unique.

So I built a Claude skill that does it for me. You point it at the folder, it reads every file on your own computer, and it hands back one Excel workbook that tells you what the tables are, how they relate and where the problems are. The CSVs never leave your machine. Claude only ever sees the workbook, which holds table and column names and counts, and zero client data.

A script reads your data. Claude reads the workbook.

The skill is two things in a folder: a SKILL.md that tells Claude how to run the analysis and how to walk you through the results, and a Python script that does the actual reading. Claude never reads the CSVs.

It works on any CSV extract: another CRM, an ERP, a SQL Server / MySQL / Postgres / Access dump, or a Salesforce Data Export from an org you are consolidating. Optional profiles (JSON files) teach it the quirks of a specific system, and it auto-detects a Salesforce Data Export and a RedTail extract without you asking.

What it works out on its own

• File basics: delimiter, encoding, ragged rows, missing or duplicate headers, dbo. and tbl prefixes.
• Primary keys: candidates verified unique and never blank across every row. A key that is not actually unique becomes a High finding.
• Foreign keys: matched by naming convention and, for GUIDs and Salesforce Ids, by value, then scored against live rows so the best resolver wins. Handles 007 vs 7, GUID braces, and 15- vs 18-character Salesforce Ids.
• Polymorphic keys: WhatId-style columns that point at several tables, and X_id + X_type pairs.
• Table classification: Entity, Child / Detail, Lookup / Picklist, Junction, Polymorphic Child, Audit / History, Staging / Backup, System / Config, each with its reason.
• Column profile: inferred type, max length, line breaks, leading zeros, populated %, distinct values, picklist candidates.

What the workbook contains

No data values anywhere in the workbook. Table names, column names, counts, percentages, types, lengths. That is it, and that is the point.

Five things, and you probably have four of them.

1. A Claude plan that runs skills. Custom skills work on Pro, Max, Team and Enterprise.
2. The Claude desktop app, on the computer that holds the client extract. Not the browser. The whole point is that the script runs locally, so Claude needs a shell on your machine. Claude Code works too if you live in a terminal; the steps below assume the desktop app.
3. Python 3. Mac has it. On Windows, grab it from python.org and tick “Add to PATH” during install. Then one package: python3 -m pip install openpyxl.
4. A client extract as CSV files in one folder. I use a standard engagement layout: 01_Source_Extract for the CSVs, 02_Analysis_Workbooks for the output. Any folder works. If what you have is a .sql backup rather than CSVs, here is how to get CSVs out of it with Navicat first.
5. The skill itself, which is the next step.

You do not need to know Python. You never edit the script. You do not need to know what a foreign key is either, although by the end of this you will.

Get the skill from GitHub.

The skill lives in a public repo: elisesfdc/source-analysis. It is MIT licensed, so use it, change it, build it into your own thing.

A skill is just a folder. This one looks like:

source-analysis/
  SKILL.md                     the instructions Claude reads
  scripts/
    source_analysis.py         the script that reads your CSVs locally
  profiles/
    _template.json             starter for a new source system
    redtail.json               RedTail CRM
    salesforce_data_export.json

1. Go to the repo’s Releases page and download source-analysis-v1.0.0.zip.
2. That zip is already shaped for the upload in Step 2: one folder at the root, named source-analysis, with SKILL.md directly inside it. Nothing to rename.
3. Unzip a second copy somewhere you will find it again. You will copy scripts/ and profiles/ out of it every time you run an analysis.

Install it in Claude.

Claude takes custom skills as a zip uploaded in Settings. The upload is per account, so each person on your team does this once.

1. In the Claude desktop app, open Settings, find the Skills section, and choose Upload skill. Pick the zip you downloaded.
2. Make sure the skill’s toggle is on. Claude reads the description line in SKILL.md to decide when to use it, so from now on “analyze these CSVs” or “what is in this export” in any chat will pull it in.

Claude Code instead? Skip the zip entirely: /plugin install source-analysis --marketplace elisesfdc/source-analysis, or clone it into ~/.claude/skills/source-analysis and type /skills to confirm. Skills installed in Claude Code do not show up in the desktop app and vice versa, so do both if you use both.

Point it at your export folder.

This is the step that makes it safe, so do not skip it.

1. Put the client’s CSVs in a folder on your computer. Mine is ClientName/01_Source_Extract/. Make an empty sibling folder for the output: ClientName/02_Analysis_Workbooks/.
2. Connect that client folder to Claude. In the desktop app, click the + next to the message box and choose Add folder, then pick the ClientName folder. This gives Claude a shell on your machine scoped to that folder, and nothing else.
3. Start a new chat and ask for the analysis.

Claude picks up the skill from the words “source analysis” and the CSV folder. The skill tells it to do two things before anything else: copy the script and profiles into a _source-analysis-tool/ folder next to your output, which brings the tool to the data, and never read, preview or upload the CSVs themselves.

If Claude asks whether a profile exists for the source system, say no unless you have built one. RedTail and Salesforce Data Export are detected automatically.

If you would rather not connect a folder at all, ask Claude for the command instead and run it yourself in Terminal:

python3 scripts/source_analysis.py "01_Source_Extract" \
    --client "Acme Wealth" --source "Junxure" \
    -o "02_Analysis_Workbooks/Acme_Wealth_Source_Data_Analysis.xlsx"

Then drag the finished workbook into the chat. Same result, and Claude still never touched the data.

Run it, and what to expect while it runs.

A typical CRM extract of 30 to 60 files and a few hundred thousand rows finishes in under a minute. Big SQL dumps take a few minutes.

The script makes two passes over every file: the first to profile columns and find primary keys, the second to test every foreign-key candidate against live rows. What you will see Claude do, in order:

1. Copy scripts/ and profiles/ into 02_Analysis_Workbooks/_source-analysis-tool/.
2. Run python3 source_analysis.py ... in the connected folder. If openpyxl is missing, it will tell you the one-line install.
3. Read the finished workbook. Only the workbook. Watch the tool calls if you are nervous; you will see it open the .xlsx and nothing from 01_Source_Extract.
4. Walk you through the tabs: Findings (High first), Engagement Intake, Source Data Analysis, Key Map, Schema.

The walkthrough is the useful part. Claude does not just dump the workbook on you; it reads the High findings out loud, tells you which key assignments it is unsure about, and asks the intake questions you will need the client to answer.

Flags you can add if a file looks odd: --delimiter tab|comma|semicolon|pipe and --encoding, both auto-detected per file so you rarely need them, --recursive for sub-folders, and --target if the destination is not Salesforce.

Reading the workbook, tab by tab.

Open the .xlsx in Excel and go left to right. Each tab answers one question.

What to do with it next.

The workbook is step one of a migration, not the migration. Three things come straight out of it.

1. 1

   The scoping call

   Take the Engagement Intake questions and the High findings to the client. You now know which tables matter, roughly how many records, and the three ugly things you need them to decide on. That is a scoping call that takes 45 minutes instead of three meetings, and from there the six phases that follow are the rest of the job.

2. 2

   The field mapping

   Filter Source Data Analysis to Included, open Schema for those tables, and map column by column to the target objects. The inferred types and max lengths tell you what custom fields to create before a single record moves, and there are faster ways to grab the target field API names than clicking through Setup.

3. 3

   A profile for next time

   If you will see this source system again, and if you do acquisitions you will, ask Claude to start a profile with --write-profile-template. Fill it with what you confirmed. Profiles hold knowledge about the system, never about the client, so they are safe to keep and reuse.

One honest limit, so you plan for it: the script will not find a foreign key that has neither a naming hint nor GUID-style values, an integer ShipVia column pointing at a Shippers table being the classic case, and it does not detect composite primary keys. Those show up as “no single-column primary key” and you add them as overrides. Worth getting right, because the keys the Key Map hands you are exactly the ones that become External Id fields on the Salesforce side. Classification is structural; confirm every row. Nothing in the workbook replaces the client signing off on scope.

The things that go wrong, and the fixes.