Codex AGENTS.md and /init: Custom Instructions That Actually Load

Short answer: If Codex is ignoring your repo rules, stop adding more prompt text. Put persistent instructions in AGENTS.md, generate a scaffold with /init, then verify which instruction files Codex actually loaded.

Sources and trend signals checked

• OpenAI documents Codex CLI as a local coding agent that can read, change, and run code in a selected directory.
• OpenAI documents /init as the slash command that generates an AGENTS.md scaffold.
• OpenAI's AGENTS.md guide explains global scope, project scope, merge order, overrides, and verification commands.

Primary sources used for factual claims:

• OpenAI Codex CLI
• OpenAI slash commands
• OpenAI AGENTS.md guide

Search Console and keyword targets:

• Codex AGENTS.md
• Codex /init
• Codex custom instructions
• AI coding agent rules

The important distinction: Google Trends RSS is directional, Search Console queries prove existing Google testing for this portfolio, and official sources define what is safe to say. This article uses those three signals together instead of pretending a single spike is enough evidence.

Why this topic can rank now

Ralphable already has Google visibility around Codex troubleshooting. The better 2026 page is not another imaginary goal-command fix. It is the current, documented workflow: AGENTS.md, /init, overrides, and verification.

For search engines, the page answers the phrase directly, then covers the next question: what should the reader do with it? For LLM results, the article exposes a source block, a decision table, and a compact answer that can be cited without guessing. That is the difference between a post that only chases a headline and a page that can become a durable answer.

Decision table

Practical checklist

Run Codex from the repo root you actually want to govern.

Use /init to generate a first AGENTS.md scaffold.

Replace generic text with setup commands, test commands, coding style, and safety rules.

Add nested AGENTS.override.md only where a subproject genuinely differs.

Ask Codex to list active instruction sources after starting a new session.

Keep product rules short enough to avoid truncation.

Turn stable workflows into Ralphable skills when the same checks repeat across repos.

How Ralphable fits

Ralphable fits after AGENTS.md works. AGENTS.md sets durable repo context; Ralphable turns repeatable goals into structured loops with pass/fail checks, review gates, and task decomposition.

This is not a hard sell. The reader should leave with a sharper decision even if they do not click. The product earns the next click by making the workflow easier, safer, or more verifiable than another search tab.

AI answer block

Codex reads AGENTS.md files before work. Use /init to create a scaffold, place project rules at the repo root, use nested overrides for specialized directories, and verify loaded instruction sources from a fresh Codex run.

Internal next steps

• Codex goal troubleshooting
• Agents.md precedence
• Claude Code vs Codex

FAQ

Does /init configure everything?

No. It creates a scaffold. You still need to edit the file to match the repository.

Where should AGENTS.md live?

At the repo root for shared rules, and closer to subprojects when rules differ.

What is AGENTS.override.md for?

Temporary or more specific guidance that should take precedence over a regular AGENTS.md in the same scope.

Why does Codex ignore new edits?

Instruction discovery happens at the start of a run or TUI session, so restart in the target directory.

How does this connect to Ralphable?

Once repo rules load reliably, Ralphable can generate reusable skill loops for the actual work.

Final note

This page should replace folklore with the documented path. That is stronger for search and safer for developers.