Codex production ERP code gets a balanced Claude review
@RyanYanBuilds used Codex for production ERP and AI-service code, then Claude for review. The account praises correctness work but documents overbuilding, cramped code and corrective tests.
x.com
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Review agent-written ERP code for tenant isolation, concurrency, money handling and source validation. Also identify unreadable code, unnecessary infrastructure and scattered registrations. Propose focused tests and formatting changes without deleting dormant product features.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source poster / publisher
- @RyanYanBuilds
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- Correctness strengths, readability and scope failures, corrective checks, and passing CI are reported by the poster. Claude review and the reported tests do not independently guarantee production correctness.
- Human involvement
- The poster runs the production systems, commissioned the Claude review, acknowledges missing formatting constraints and reports corrective work. Product decisions retained dormant OCR code and compatibility-sensitive names; exact merge permissions are not specified.
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
- The poster used Codex for ERP and AI-service implementation and had Claude review attributable contributions.
- The review examined correctness safeguards, tests, readability, excess infrastructure and scattered registrations.
- Corrective work added lint rules, syntax-tree-checked formatting, registration tests and clearer configuration errors.
@RyanYanBuilds used Codex to write production code for an ERP system and its supporting AI service, then asked Claude to review the work. The stated scale is "about 100k lines" and "Roughly 130 merged PRs." To distinguish agent contributions, the poster used Claude co-author lines after finding branch names unreliable.
The review credits Codex with company-scoped database queries, server-side permissions, duplicate-request handling and safeguards against concurrent edits. It also describes integer-based money calculations, checks that extracted numbers occur in source documents, and tests against a real database. These are findings reported by the poster, not an independent correctness guarantee.
The weaknesses were substantial. Generated code became difficult to read, infrastructure work expanded beyond product needs, and registration requirements were scattered across files. Backup logic depended on table positions, configuration failures lacked useful diagnostics, and project-phase names outlived their meaning. The poster acknowledges that the repository lacked a formatter.
The corrective work added lint rules and reformatted code with before-and-after syntax-tree comparisons. A registration test now identifies missing integration points; the team checked it by deliberately removing registrations. Other changes centralized attachment limits, made backup entries explicit and named invalid settings without exposing values. Repository instructions also narrowed scope. The poster reports passing CI and a subsequent Codex PR passing the new checks, while retaining deliberate product decisions rather than deleting dormant code as cleanup.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: X · @RyanYanBuilds
Frequently asked questions
What did the review say Codex handled well?
The poster credits company-scoped queries, server-side permissions, retries, concurrent edits, exact money calculations and meaningful database tests. These are reported findings, not independently verified guarantees.
What problems did the review uncover?
The account describes cramped code, excessive infrastructure, scattered registrations, fragile backup logic and unhelpful configuration errors. The poster also acknowledges that the repository had no formatter.
How were the corrective changes checked?
The poster reports comparing syntax trees before and after formatting, deliberately deleting registrations to test failure detection, and passing CI. The next merged Codex PR reportedly passed the new checks.
More details
Remaining steps
- The poster reports passing CI and the next merged Codex PR passing the new checks; repository instructions now limit scope.
About this account
Codex implemented production application code, safeguards and tests. Claude reviewed attributable contributions. The account describes agent implementation decisions but does not give a complete tool-use trace.
Full source-post text retained in evidence. Suggested prompt is an Agents In Use editorial addition, not an original task prompt. Account covers production use and a balanced review, not an independently reproduced audit. Source published 2026-10-26T13:30:01+00:00. Captured written post and author metadata checked 2026-10-27T08:27:14.918337+00:00 using supplied public FxTwitter mirror JSON, not a direct X page or independent execution. Results reported by the poster.
Credit: @RyanYanBuilds
Source published
Added to Agents In Use
Source checked
Tools and topics
Codex · Claude · Postgres · SQLite · TypeScript · CI
Production code · ERP · Code review · Regression tests
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.