Claude struggled to follow file-reading instructions before coding
In a real coding setup, Claude did not reliably follow an instruction to read a specific file before writing code. The project was a monorepo with separate client and server . Claude and Codex were both used in the workflow, with one tool planning or implementing and the other reviewing.
The instruction files for Claude and Codex were kept identical and short at first, mainly listing strict rules such as never doing and checking a specific tool before coding. As the instruction files grew to about 200 lines, the rules were split into smaller files because the client and server had different needs and reading every time would waste tokens. A main rules file was set up to point Claude toward separate client and server rule files.
The problem was that Claude could not be trusted to follow those linked file-reading steps consistently, so the main instruction file was reduced back to one critical rule: never perform .
Key points
- Claude did not reliably read required rule files before writing code.
- Claude and Codex were used together, with one implementing and the other reviewing.
- The project used a monorepo with separate client and server .
- Rules were split because the instruction files had grown to about 200 lines.
- Critical rules should stay in the top-level instruction file, not only in linked files.