Automate marketing reporting with Codex

The starting point

• Two representative anonymised exports, including missing data
• A dictionary of columns, currencies, dates and calculation rules
• A manually calculated reference for a small sample

1. Define a data contract

Specify required columns and types, row granularity, duplicate keys and date rules. Have the program clearly reject incompatible files. Missing values should not silently become zero: define a documented rule for each field.

2. Develop the script and checks

Ask Codex for reading, validation, aggregation, calculation and export in your approved environment. Disallow unnecessary network access and source-file modification. Supply reference results and error cases. Review the code and run tests before recurring use.

3. Add commentary

Give GPT or Claude the validated tables and business context. If evaluating DeepSeek via API, test the same format, quality and data conditions. The model describes observations; it does not reinvent totals or silently repair rejected input.

4. Plan operation

Keep execution logs, script versions and input identifiers. Define handling for changed exports or unavailable services. Schedule only with an owner, duplicate protection and a stop mechanism. Document manual recovery.

A practical example

Build a reporting script from the attached schema and examples. Do not modify source files or access the network. Validate columns, types, dates and duplicates. Aggregate counts before calculating ratios; represent zero denominators as not calculable. Provide a results export, validation log and tests for valid input, missing columns, duplicates and invalid values. Compare against the small reference sample before considering it ready.

How to assess progress

• A script that runs is not necessarily correct.
• Never put secrets or private data in a public repository.
• Assign maintenance and account for export-format changes.