Keep only the CSV columns you need

Selection is positional, even when it starts with a name

A delimited table is a sequence of rows, and each row is a sequence of fields. The optional header gives those positions names, but the data underneath still belongs to positions. Selecting Email therefore means: find the one header field whose trimmed text is exactly Email, remember its zero-based internal position, and copy that position from every later row.

That distinction matters when headings repeat. Given:

Name,Email,Email
Ada,home@example.test,work@example.test

the word Email does not identify one column. Guessing the first would create a valid-looking file with the wrong address, so the tool refuses the ambiguity. #2 selects the first email column and #3 selects the second. Positions shown to people are one-based because #1 is the first visible spreadsheet column.

A blank keep-list is useful too: it selects every position. This lets the tool normalise a separator or verify a file before the visitor narrows it.

The parser cannot split on commas

The separator is structure only when it is outside a quoted field. These are three fields, not four:

Name,City,Note
Ada,London,"writes, tests"

A quoted field may also span physical lines, and a literal quote inside it is written twice. The parser therefore walks characters with one piece of state: whether it is inside quotes. Outside quotes, a separator ends a field and a line break ends a record. Inside quotes, both are text. Two quotes in a row add one quote to the current value without leaving that state.

Automatic separator detection applies the same rule to the first complete record. It counts comma, semicolon, tab and pipe candidates only outside quotes. An explicit choice remains available because no guess can recover a separator that does not appear in a one-column header.

A worked selection

Start with this four-column file:

Name,Email,City,Note
Ada,ada@example.test,London,"writes, tests"
Grace,grace@example.test,New York,"line one
line two"

Enter the following keep-list:

Email
Name

The result contains 2 data rows, keeps 2 of 4 columns, and puts Email before Name:

Email,Name
ada@example.test,Ada
grace@example.test,Grace

The comma and embedded line break in the discarded Note column are still parsed correctly; neither can create a phantom row or shift City under another heading. Those exact inputs and the exact CRLF output bytes are asserted in the tool’s tests.

What changes in the downloaded bytes

The selected cell values remain text. They are not parsed as dates or numbers, so a code such as 007 stays 007. A short row gets an empty value for a selected position it does not contain, rather than allowing the next field to slide sideways.

The physical representation can change. Records are written with CRLF endings, quotes are added only where the chosen output separator or a line break requires them, and unnecessary source quotes may disappear. A UTF-8 byte-order mark is restored only when the input carried one. These changes preserve the selected table, not byte-for-byte formatting around it.

Use CSV to JSON when the next system needs named objects, or JSON to CSV when a structured response needs to become a flat table. The guide on rows, records and columns explains why those shape decisions cannot be inferred from punctuation alone.