Model applies its own judgement instead of literally executing a data-transformation prompt

What happened

I asked Claude Code to generate a WooCommerce product CSV and a rules JSON from two source text files (a GloriaFood menu dump and an add-ons dump), following a given JSON template. The model kept swapping in its own judgement instead of doing the transformation literally:

• It changed option prices from the source to other values it thought were right.
• It added features nobody asked for, like wizard steps and delivery-only visibility rules.
• It made up min/required values that aren't in the source.
• It spent many turns probing the live site's REST API instead of producing the files.
• After I corrected it, it misunderstood the correction and rewrote the output files again without being asked.

Expected

A literal, fast A-to-B transformation of the source into the template. Where something is missing from the source, one line flagging it. No extra "improvements" and no side investigations.

Repro

Give Claude Code a source data file plus a target template and ask for a literal conversion. It adds its own decisions, reports them as "decisions taken", and burns time on unrequested exploration.

Environment

Claude Code on Windows 11, model Opus 5 (1M context), background session.