Use Codex to catch an S-Corp tax field mapping error
John Sambrook prepared S-Corp tax materials with Claude Code and Codex working in different roles. QuickBooks data and bank statements went into a private Git repository; Claude wrote Python scripts to compute forms, while Codex reviewed those scripts using a model configured for deeper reasoning.
common-sense.com
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Review these S-Corp scripts against source financial JSON. Trace Schedule L outputs to their input fields, list discrepancies by severity, and review revisions after fixes.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- John Sambrook
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- Codex flagged three issues, including the Schedule L line-eighteen mapping error; Claude repaired findings through repeated review under Sambrook's supervision.
- Human involvement
- Sambrook directed the process, checked results and made tax decisions himself.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Put financial records and calculation scripts in a private repository.
- Have a second agent trace computed fields back to source data.
- Fix findings, repeat review and personally check generated figures.
John Sambrook prepared S-Corp tax materials with Claude Code and Codex working in different roles. QuickBooks data and bank statements went into a private Git repository; Claude wrote Python scripts to compute forms, while Codex reviewed those scripts using a model configured for deeper reasoning.
One review of the 1120S scripts produced three findings ordered by severity. The most serious concerned Schedule L: the ending value for line eighteen had been derived from total liabilities instead of other current liabilities. Codex found the mismatch by following data from source JSON into the generated CSV. The review examined the meaning of a field, rather than whether a program merely ran.
Sambrook describes a repeated cycle: Codex identifies an issue, Claude repairs it, and Codex reviews the revision. He directs that process, checks the resulting figures, and makes decisions. The article reports preparation and correction work; it does not establish that agreement between the agents substitutes for his review.
Keeping statements beside the calculations also helped with related bookkeeping. When Sambrook asked to reconcile a business credit card, Claude identified the supporting statement's full local path and explained what data was needed.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: common-sense.com · John Sambrook
Frequently asked questions
What made the serious finding concrete?
The generated value came from total liabilities rather than the intended other-current-liabilities source field.
Why use a second agent?
Sambrook assigned Codex to review and Claude to repair, repeating the cycle when findings remained.
Who accepted the tax results?
Sambrook says he directed the work, checked results and made the decisions.
More details
About this account
Codex traced source JSON into CSV outputs and reviewed Claude's calculation scripts.
Firsthand source opened and read. Suggested prompt written by us from the reported workflow; no original task prompt is quoted.
Credit: John Sambrook
Source published
Added to Agents In Use
Source checked
Tools and topics
Codex · Claude Code Opus 4.6 · QuickBooks · Python
S-Corp taxes · Schedule L · Cross-agent review
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.