Four recurring tasks, packaged as installable Codex agents

The problem and the idea

I repeatedly handle four kinds of work: refining prompts, examining arguments, checking facts, and auditing investment theses. Each time, I used to explain the task boundaries, evidence standards, and restrictions again. The prompts grew longer, yet behavior still drifted between conversations.

The idea I wanted to test was simple: could narrowly scoped custom agents, delivered with an installer, documentation, and regression checks, be more consistent and easier to maintain than a few saved prompts?

The Codex documentation describes personal or project-level custom agents defined in separate TOML files, each with a name, description, and instructions. This repository is my personal configuration package built on that mechanism. It is not an official OpenAI product or agent collection.

Constraints and decisions

I divided the work into four roles according to the kind of judgment involved:

• _mantou clarifies and refines prompts without carrying out the tasks inside them.
• _manuel uses a critical-thinking framework to examine questions, evidence, and inferences.
• _factbot separates claims, evidence, inferences, and opinions, checking source independence and uncertainty.
• _invest supports long-term investment research and thesis review, without executing securities trades.

The boundary around public material is part of the package. Personal Obsidian paths and case indexes live in ignored local files. The repository contains no private notes, actual holdings, credentials, or source texts from reference books.

What I shipped and checked

The current public version is v0.1.1 . Source code and full instructions are in the GitHub repository . This stage delivered:

• Configurations, role descriptions, and invocation examples for four agents;
• A PowerShell installer with preview and forced-update options;
• Isolated installation regression tests that do not write to the real Codex directory;
• Checks for configuration syntax, file integrity, privacy boundaries, and installation results;
• Documentation covering compatibility, contributions, security, changes, and releases;
• A versioned GitHub Release with change notes.

Both repository validation and isolated installation from the release archive passed. Version v0.1.1 also fixed validation on Windows systems where python is not available: the scripts try the standard py -3 launcher.

Limitations

• Automated tests establish that configuration and installation behave as expected. They do not guarantee correct answers.
• Different models, reasoning settings, Codex versions, and permissions can produce different results.
• Fact-checking and investment research still require checking sources; an agent’s name is no guarantee of reliability.
• _invest supports research and review only. It cannot trade automatically.
• The official custom-agent format may evolve, so compatibility needs ongoing attention.

What comes next

This stage turned a few prompts I often use into a versioned package that can be installed, checked, updated, and tested again. Engineering completeness is only the first layer of evidence. Its value still depends on real task performance.

Before adding a fifth agent, I will document real uses of the existing four: the task, where the first answer failed, the corrections I made, and whether a repeated task required less work. Those records will determine whether to add roles or more systematic behavioral evaluations.

For the platform mechanism and configuration boundaries, refer to the official Codex documentation .