data-transform

Data Max — Data Transform

Turn a review into files. Given a <filename>-review.md (the spec) and the original source file, produce the cleaned warehouse-ready CSV(s) — one file, or N files when the review decided different grains belong apart. The work product is an ad-hoc, rerunnable script under tools/ plus the output data, so the next export can be cleaned the same way without re-deciding anything.

This skill executes; it does not decide. The why behind every change — money→mills, ISO dates, unpivot, tenant column, single-vs-multiple — lives in the data-review skill. Here you assume the report already justified each change and you faithfully implement its "🎯 Target CSV shape".

Required input — do not skip

You need the review report (<filename>-review.md) and the original file.

digraph gate {
  "Have the review report?" [shape=diamond];
  "Run the data-review skill first" [shape=box];
  "Proceed: implement its Target CSV shape" [shape=box];
  "Have the review report?" -> "Run the data-review skill first" [label="no"];
  "Have the review report?" -> "Proceed: implement its Target CSV shape" [label="yes"];
}

If there is no report, stop and run the data-review skill first — that is how the transformation is specified. Don't reverse-engineer a spec by transforming on instinct: the review is where isolation, grain, and tier decisions were made with the user. The only exception is a tiny, unambiguous one-liner the user states explicitly ("just rename the headers to snake_case"); even then, prefer a quick review so the change is documented.

Workflow

1. Read the report and the original file. Anchor on the report's 🎯 Target CSV shape and its single-vs-multiple decision — that is your output contract. Note every 🔴/🟡 item the user accepted. If the user said they were constrained, honor the 🪜 Minimal viable change instead of the full set.

2. Plan the passes. Decide how many scripts and how many output files:

   • Group all row-level fixes that share one read-write pass — strip junk + rename + money→mills + ISO dates + tenant column — into one cohesive script. Don't chain micro-scripts that each re-parse the file.
   • A reshape that changes the grain (unpivot wide→long, or splitting file A into N grain-specific files) is its own concern — separate script and/or the N output files the review named.