A music producer built working audio plugins for Logic Pro using ChatGPT and OpenAI's Codex desktop app, and then put them through a blind listening test against commercial mastering tools. The results cut both ways. A soft clipper assembled in roughly ten minutes placed second out of seven processors, ahead of several paid products. A more ambitious mastering plugin, designed and coded in about four minutes of AI working time, landed in sixth place. The gap between those two outcomes is the real lesson: the AI handled the code, but the sound still depended on the producer's ears and audio knowledge.

The setup: a $20 subscription and a desktop agent

The build ran on a standard ChatGPT Plus subscription, cited at about $20 per month. Inside the OpenAI desktop app, the engine was switched from regular ChatGPT to Codex, OpenAI's local coding agent. Codex can operate directly on the computer: it creates and edits files, runs scripts and build tools, and compiles software without asking for permission at every step. That matters for audio work, because a plugin is not a chat answer. It is a compiled binary that has to be installed where a digital audio workstation (DAW) such as Logic Pro can load it.

Two settings shaped the cost of the session:

  • Model: GPT-6 Astra, described as OpenAI's top model for complex engineering tasks. Think of model tiers as an organization chart, where smaller models handle junior tasks and flagship models act like executives who can design architecture. Using the flagship for trivial work wastes quota.
  • Reasoning effort: set to Light. Higher reasoning tiers consume far more of the weekly token budget, and the Ultra setting also prompts for system permissions as it explores more autonomously. Light reasoning proved sufficient for the prototype.

Anthropic's Claude Code, running Claude Opus 5.5 in High Mode, looked like the stronger coding option at the moment, but Codex was chosen for the walkthrough because of its mainstream accessibility.

Prompting from two reference images instead of a spec

Rather than describe a mastering chain from memory, the producer gave the AI pictures. A screenshot of Musik Hack's Master Plan 2, a commercial mastering plugin, went into the prompt with instructions to identify its mastering components and reinvent them under a new name and architecture. The prompt explicitly required every parameter name to be reimagined, said the result must not be a visual or functional look-alike, and told Codex not to pull code from any existing repository on the machine. The goal was to borrow the concept of what a mastering plugin does, not its intellectual property.

For the look, a product photo of the Drawmer MC3.1 monitor controller, a piece of studio hardware with tactile vintage-style knobs, was pasted in with a request to design an original interface using similar knobs, texture and colors. Combining a functional reference from software with a visual reference from hardware produced something that resembled neither.

Two reference pictures beside a monitor where a new charcoal plugin interface with knobs and amber meters takes shape

▲ Software and hardware references feeding a new design

In about one minute and twenty seconds, Codex returned an interface mockup and a design document. The concept, working-titled Plan B, used sculpted dark knobs, charcoal metal faceplates, ivory text and warm amber meters, with controls grouped into four blocks:

Block Controls
Shape Multiband compression, frequency-balancing EQ, clarity enhancement, resonance control
Texture Tape warmth, saturation, cohesion, high-frequency harmonic generation
Space Stereo field management, width, mono audition
Finish Output level, density compression, loudness limiting, true-peak ceiling

A separate Detail tab was specified so deeper parameters would not clutter the main panel. Codex was clear that this was only a plan: no audio processing code existed yet.

Where it stalled, and how the prompt fixed it

The first real obstacle was not technical. When the task started, Codex checked the account's weekly usage, found it at 88 percent, and stopped because an AGENTS.md file (a plain-text configuration file that sets rules for the agent) defined a cutoff at 80 percent. The agent's message said it had halted the work to avoid runaway consumption.

The fix was a one-line manual override: proceed, but spend no more than 5 percent of the weekly allowance on this step. That kind of guardrail is worth copying. Autonomous agents can burn through a quota on experiments, and a written cap gives them a reason to stop and ask.

With the override in place, the next prompt asked Codex to write the DSP library (digital signal processing, the math that actually changes the audio) from scratch, plus a small room reverb inside the Space section. That stage took two minutes and 38 seconds and used about 1 percent of the weekly budget. The resulting core included a six-band multiband network, a room reverb with a 0 to 25 percent mix range that is bypassed by default, true-peak overshoot protection and native oversampling. Total generation time for design and DSP came to roughly four minutes, work that historically took software engineers nine months to several years.

The final prompt was plain English: install the VST and AU formats so the plugin can be used in Logic Pro. Codex configured the local build tools and compiled native macOS binaries, with build runs logged at between one minute 15 seconds and two minutes 24 seconds.

The first build had real audio flaws

Loading Plan B in Logic Pro showed how far a working prototype is from a finished product. The problems fell into two groups.

Audio engineering problems the AI could not judge on its own:

  • The Room Mix control blended reverb as a direct crossfade instead of a parallel send, so turning up ambience lowered the dry signal and sapped the mix's punch. The prompt should have specified parallel routing from the start.
  • The Drive saturation sounded generic, without notable warmth.
  • Pushing the Finish-section limiter, with the ceiling set to -0.3 dBFS, produced a crunchy, flattened sound that lacked the smoothness and headroom of commercial limiters.
  • A control named Luster produced almost no audible change, and its vague name told users nothing about what it did.

Typical AI build quirks:

  • Garbled or extraneous label text on the interface.
  • Two buttons labeled TX and KX pulled from a design rule in an internal project folder, despite the instruction not to copy existing code.
  • The bypass control rendered as a 0 to 100 percent dial instead of an on/off switch.
  • The compiled GUI looked nothing like the approved mockup.

The last issue was resolved with another short prompt in ChatGPT, stating that the result did not match the reference and asking for a rebuild of the layout according to the mockup. The rebuild finished in under two minutes and used about 2 percent of the weekly budget. Compared with Logic Pro's bare Controls view, which exposes every parameter as a raw vertical slider, the structured panel made the plugin far easier to understand, a reminder that a plugin without a clear visual hierarchy is just a pile of sliders.

From the first prompt through the rebuild to the blind test No Yes Prompt with tworeference images If the usage cap trips,override with a 5 percent limit Design document,then DSP library Build and installVST and AU Load in Logic Proand listen Matches themockup? Blind test againstcommercial tools
▲ From the first prompt through the rebuild to the blind test

The blind test that settled it

Sound quality was judged with HOFA 4U+ BlindTest, a Logic Pro utility that hides which plugin is on which slot behind dice icons and lets the listener rate and annotate each one. Seven processors ran on the same hip-hop instrumental: Plan B, the ten-minute ChatGPT soft clipper, the producer's own existing AI mastering tool (version 13), Master Plan v1.5 using its Quick Master preset, Master Plan 2, the Apogee Limiter and an open-source test limiter.

Monitor showing a blind listening test with seven hidden slots, star ratings and notes, listener wearing headphones

▲ Seven-plugin blind listening test

Notes typed during the rounds sorted the slots into three tiers: dark and too soft at the bottom, clean but "too safe" transient-focused options in the middle, and loud, punchy contenders at the top. One slot stood out immediately with the note "wide, clear, full dense, punchy, louder." Another was marked "overly crunchy, distorted, too soft." A direct A/B between the top two described the gap as substantial, with the leader showing more stereo depth, width and high-end detail.

With the blindfold removed, the ranking was:

Rank Plugin Listening note
1 Producer's existing AI mastering tool (v13) Fullness, density, punch, wide stereo image
2 ChatGPT-built soft clipper (about 10 minutes to create, free) Clear, solid transients, slightly darker
3 Master Plan 2 Loud but overly crunchy
4 Master Plan 1.5 Too crunchy
5 Apogee Limiter Transient-focused but too safe and quiet
6 Plan B (new build) Overly crunchy, distorted, too soft
7 Open-source test limiter Lowest

These rankings reflect one listener's preferences on one track, with an admitted lean toward strong transients for beat-driven music. A softer, darker processor might suit a nostalgic, smooth production even if it fails on an energetic beat. Still, the picture is consistent: a tiny ChatGPT-built clipper outranked commercial tools, one of which is described as having taken three years to develop, while a far more complex AI-built plugin finished near the bottom. Generating a hundred mastering plugins might take around 400 minutes at this pace, but speed says nothing about sound.

What to take from this

The argument here is that coding is no longer the bottleneck in audio tool development. Writing DSP and GUI code by hand once took months or years; here it took minutes and a few percent of a weekly AI budget. What did not get easier is knowing what sounds good, how to route a reverb, which saturation is musical and when a limiter is crushing the transients. That judgment produced the second-place clipper and exposed the flaws in Plan B.

If you want to build your own tool, in music or any other field, the pattern transfers:

  1. Give the model reference images of existing tools to extract functional blocks, and a separate visual reference for the look.
  2. State clearly what must not be copied: names, trademarks, source code.
  3. Start on a light reasoning setting and set a usage cap in an agent configuration file before letting it run.
  4. Be specific about control types and signal flow up front, such as a toggle for bypass and parallel routing for ambience.
  5. When the output misses the reference, say so plainly and ask for a rebuild.
  6. Test the result blind against the best alternatives you can get, including trial versions, with seven to ten contenders rather than a single one-on-one comparison, so confirmation bias cannot flatter your own creation.

The tooling will keep changing, with model names and versions turning over quickly. The skill that holds its value is the domain expertise that tells the AI what to build and the critical judgment that tells you whether it worked.