Just as the reader maps rows to a header, the writer should turn name-to-value records back into rows under a header. Today you add that, the write-side mirror of header mode, so keyed data becomes a proper CSV with a header row.
Write a header row and then each map's values in header order, filling missing keys with the empty string.
Header mode made reading produce name-to-value maps, and a symmetric writer completes the pattern by consuming them. You give it the column order as an explicit header, because a map has no inherent order and the columns must line up the same way in every row, then it writes the header as the first record and each subsequent row by looking up each header name in that row’s map. The result is a well-formed CSV whose first line names the columns, which is what most consumers expect.
The one rule to fix is what happens when a row is missing a key the header names: it is written as an empty field, so every output row has the same number of columns as the header regardless of which keys each map happened to contain. This mirrors the reader’s short-row rule and keeps the table rectangular. Build the header and rows into a plain list of records and hand them to the writer you already have, so all the quoting, escaping, and terminator logic is reused unchanged. That reuse is the sign the library’s pieces compose cleanly, which the capstone is about to lean on completely.
// WriteMaps(headers []string, rows []map[string]string) string:// first record is the headers themselves// for each row, emit headers[i] -> row[headers[i]] (missing key -> "")// build a [][]string and reuse Write, so quoting and terminators come for free