Build an internal tool with Codex: from bottleneck to pilot

Choose a small, frequent bottleneck

Our example is a service business that combines three spreadsheet exports each week. Staff currently rename columns, remove duplicate records and check totals. The pilot turns this into repeatable processing with an output file and an understandable error list. It initially replaces one bounded step, not the entire business process.

Before building, describe the inputs and intended transformation: what should be preserved, combined or deliberately rejected? An existing manual process often provides better criteria than a long feature wish list.

Keep changes traceable

Codex supports Git worktrees for separate working states. For the pilot, use a separate working state with test files before connecting production folders. Keep the baseline, change and tested result together so a faulty modification can be reverted deliberately.

Business rules belong with the project

AGENTS.md can provide Codex with project-specific working instructions. In our spreadsheet example, include the required checks, permitted test data and business rules: the same customer number does not automatically mean the same order, empty amounts are not zero, and unknown columns must not silently disappear.

Also describe which actions require human approval. Written rules supplement technical permissions; they do not replace bounded access to files, accounts and destination systems.

Test the failures that happen in daily work

The regular export is only one test case. Add missing required columns, duplicate order numbers, different date formats and an interrupted write. The output must either be correct or stop clearly. A file saved successfully can still be wrong in business terms.

• Compare totals and record counts with a manually verified reference result.
• Preserve original files and identify affected rows in errors.
• Review the change as a diff; automated review does not replace domain acceptance.

A usable tool needs an operational handover

Handover includes startup instructions, known limitations, sample files, an owner and a fallback. Ask a future team member to operate the tool using the instructions. If they need the original developer to complete the task, the handover is not finished.

A Codex Business Implementation is appropriate when individual tools should become a repeatable development and review process. The AI System Check first helps determine whether your bottleneck needs a new tool, an integration or better use of existing systems.