How Good Is GPT-5 Codex? I Built an App

Key Concepts

• Codeex CLI (GPT5 Codeex): A coding tool integrated within an IDE (VS Code in this case) used to automate the creation of a local transcription application.
• PRD (Product Requirements Document): A document outlining the features and specifications of the application to be built.
• MLX Framework: A framework used for loading the speech-to-text model (Whisper) and potentially LLMs for grammatical correction.
• Whisper Base Model: A speech-to-text model used for transcribing audio.
• Agentic Coding: The ability of a coding tool to work independently on complex tasks, iterating on its implementation and fixing errors.
• Hotkeys: Keyboard shortcuts used to start and stop the transcription process.
• Virtual Environment: An isolated environment for running the Python application with specific dependencies.
• LLM (Large Language Model): Potentially used for grammatical correction of the transcribed text.

Recreating a Local Transcription App with Codeex CLI

Project Goal

The goal is to recreate a local transcription application that converts audio into text in any text box, using Codeex CLI and following a provided PRD. The application should:

• Run on Apple Silicon.
• Use a Python backend.
• Utilize the MLX framework.
• Employ the Whisper base model for speech-to-text.
• Incorporate hotkeys for starting and stopping transcription.
• Display the model being used in the menu bar.
• Track and display transcription statistics (words transcribed, keystrokes saved, total duration saved).

Initial Setup and PRD

• The project is initiated using Codeex CLI integrated with VS Code.
• The PRD is provided to Codeex CLI as a markdown file (PRD.md).
• GPT5 Codeex (specifically fine-tuned for coding) is used with "high" settings.
• The system is named "Verbby" (a working title).
• The presenter provides code snippets within the PRD to guide Codeex CLI.

Codeex CLI's Initial Implementation (9 minutes)

• Codeex CLI analyzes the PRD and creates a task list:
  • Understanding requirements and designing module layout.
  • Implementing transcription recording, hotkey functionality, and menu bar integration.
  • Adding supporting assets.
• Codeex CLI generates initial code and creates a project folder.

Testing and Debugging (First Iteration)

• A virtual environment ("codeex_word_env") is created using Python 3.10.
• Requirements are installed from the generated requirements.txt file.
• The application is run, but initially, the hotkeys do not work, and an error occurs during transcription.
• The error is related to the loader path, requiring documentation lookup.

Addressing the Documentation Issue (1 minute 45 seconds)

• Codeex CLI is informed about the error and attempts to fix it.
• The presenter notes a common issue with coding agents: pinning specific package versions, which can cause conflicts.
• Codeex CLI cannot access the internet by default, preventing it from directly accessing documentation.
• Codeex CLI updates the loader path based on the package's runtime exposure.

Second Iteration and Partial Success

• The hotkeys now trigger something, but the transcription still fails with the same error.
• The presenter reiterates the need for Codeex CLI to consult the documentation.
• The presenter points out that Codeex CLI is not using fixed package versions.

Resolving the Transcription Error

• After further debugging, Codeex CLI successfully downloads the model.
• Transcription now works, but only when triggered from the menu bar (start/stop transcription options).
• The hotkeys are still not triggering the transcription pipeline.

Hotkey Fix (6-7 minutes)

• The presenter informs Codeex CLI about the hotkey issue.
• Codeex CLI implements a fix, changing the hotkeys to Command+Shift.
• The new hotkeys now trigger the transcription, and audio feedback is heard.

Adding Model Display to Menu Bar

• The presenter requests Codeex CLI to display the model being used for transcription in the menu bar.
• Codeex CLI implements this feature, showing "Whisper Small MLX" in the menu bar.

Implementing Transcription Statistics

• The presenter asks Codeex CLI to add statistics tracking:
  • Words transcribed.
  • Keystrokes saved.
  • Total duration saved.
• The statistics should persist across different sessions.
• Codeex CLI implements the statistics feature, displaying the data in the menu bar.

Verification and Observation

• The presenter verifies the accuracy of the displayed statistics.
• The presenter observes that the model seems to be unloaded and reloaded after each transcription.

GPT5 Codeex Performance and Conclusion

• The presenter expresses being impressed with GPT5 Codeex on "high" settings.
• The tool is considered a "really good agentic coding tool."
• The presenter estimates that a usable transcription application was created within approximately 30 minutes.
• The presenter plans to add an LLM for grammatical correction and create a standalone version of the application in future videos.
• The presenter encourages viewers to sign up for notifications regarding the transcription app project.

Notable Quotes

• "This is actually realtime transcription. It's using whisper base model. So, it's a relatively small model, but still it's pretty accurate."
• "GPT5 codeex on high setting is a really good agentic coding tool and I'll highly recommend to test it out."

Data and Statistics

• Initial implementation took 9 minutes.
• Fixing the first error took 1 minute 45 seconds.
• The entire process of creating a usable transcription app took approximately 30 minutes.
• GPT5 can generate 90% less tokens for the 10th percentile of the easiest tasks.
• GPT5 can generate up to 100% more tokens for the last 10 percentile in terms of complexity.

Synthesis/Conclusion

The video demonstrates the use of Codeex CLI (GPT5 Codeex) to rapidly prototype a functional local transcription application. While initial errors and debugging were required, the tool successfully implemented the core features outlined in the PRD, including hotkey control, transcription, model display, and statistics tracking. The presenter highlights the potential of GPT5 Codeex as an agentic coding tool for accelerating software development. The key takeaway is that Codeex CLI, particularly with the GPT5 Codeex model, can significantly reduce the time and effort required to build real-world applications, even with limited internet access.