Build an AI coding editorial agent with Codex without generating hollow content

An editorial agent is not useful because it writes mediocre text faster.

If the prompt is only “write an article about Cursor”, you will get exactly what you wanted to avoid: a soft intro, three generic points, an interchangeable ending, and prose that smells generated.

The better use is harder: make it filter topics, verify sources, frame the tension, then write.

1. Start with the editorial contract

Before the agent, write the rules that do not move:

• audience: senior developers, technical founders, builders;
• tone: direct, concrete, skeptical;
• forbidden patterns: empty hype, artificial SEO, “in this article”;
• requirement: cite sources when the topic is recent;
• output: Markdown compatible with the Astro collection.

In Codex, broad repo rules go in AGENTS.md. The detailed workflow goes in a skill.

2. Make a skill, not a throwaway prompt

A skill named ai-coding-tutorial should say when it triggers and what it produces.

Minimum header:

---
name: ai-coding-tutorial
description: Use when creating /ai-coding tutorials about Codex, Claude Code, Copilot, Cursor, MCP, agents, hooks, context, or AI coding workflows.
---

Then add the workflow:

1. read AGENTS.md;
2. inspect src/content.config.ts;
3. read one existing tutorial;
4. verify recent facts;
5. write FR + EN;
6. run bun run build.

This prevents two common mistakes: writing to the wrong collection and forgetting the English mirror.

3. Give it a refusal grid

A weak editorial agent accepts every topic. A good one can say: “not solid enough”.

Useful criteria:

• does the topic actually change a developer workflow?
• is the source official or primary?
• can the piece include a concrete decision?
• is there a cost, risk, limit, or review burden?

If not, keep it as a newsletter note, not a tutorial.

4. Split research and writing

For a recent topic, ask first:

Use the ai-coding-tutorial skill to propose 5 angles.
Do not draft yet.
For each angle, give the official source and the bullshit risk.

Then:

Draft angle 2 as a /ai-coding tutorial.
Keep a field-level tone, not marketing.
Add the English version.

You keep control of the editorial line, and the agent works on a validated topic.

5. The format that works

For /ai-coding, aim for a practical tutorial:

• concrete tension in the opening;
• short context;
• decision grid;
• step-by-step workflow;
• traps to avoid;
• sources.

The reader should leave with a way to work, not only an opinion.

Sources

• OpenAI Codex - Agent Skills
• OpenAI Codex - Custom agents and subagents