Since I’ve been working with Muse and Codex (aka ChatGPT Work – confusing branding!) in parallel and manually passing tasks back and forth between the two, it’s been a beast to have to juggle metered usage limits of each tool. Never mind figuring out the best uses of each one’s architectures, as they are rather different on their mysterious “back end”… and then comparing quality across models for creative generations.

As I wait patiently for my 5 hour usage limit to reset, to see if I can put this all into action, I’m pondering and I think it might be possible to establish an orchestration which uses both tools in concert as cheaply as possible (I’m hoping) with the various prior work streams I have developed in each, and pitching to each one’s strengths….

For example, Muse’s creative text generations and image results are in my opinion 2-3 steps behind how good they’ve gotten in ChatGPT running the newest image model there. I won’t say that ChatGPT crushes images every single time, but the moderate success rate is high, and when it does crush them, it crushes them into diamond dust… whatever that means.

So there’s no point in asking for anything but secondary thematic filler images with the state of things in Muse. And quality of the stories that it develops textually are so far nothing that I feel is good enough quality – or bad enough quality, but in a good enough or noteworthy way – to warrant inclusion in a published volume of the Lore Books. But I’ve had writing developed as part of my ChatGPT Work pipeline runs using Astra Light which have blown me away, and in at least some cases don’t seem to require any editing at all.

So there’s no point in asking Muse for manuscript text generations, but perhaps it could help with more mundane productions like product descriptions for the bookstore page on Payhip. (To be determined.)

And it also did a terrible job in assembling a .docx file for import into Vellum. ChatGPT Work can be cajoled to do an almost perfect version there too. (Needs some tinkering still.)

But ChatGPT Work/Codex eats of up usage fast. I started with 75% usage in a 5 hr window on Plus plan, and did not end with a finished docx file on a complete end-to-end run using just the pipeline and skill documents (never mind my generative control surface). Ostensibly, all that run really did was lightly reference a corpus (which already has a JSON file containing extracted contents of EPUB corpus), write 2000 words in a loosely connected setting, generate 8 body images, and then about 5 or so revisions to text treatment for a cover onto one of the existing body images. Then it mapped the image assets where they go to the text, and metered out before finalizing its docx assembly.

Granted, I’m sure my pipeline is not adequately optimized, and I’ve already experimented with delegating via browser access some text & image gen work into Chatgpt Plus regular chat threads, and then pulling the results back into Work for manipulation… This way the regular chats use their own different less intensive metering, and Work usage is somewhat reduced.

On the other hand, because of Muse’s seemingly (anecdotally, I have nothing to prove this) looser usage restrictions than ChatGPT Plus, I’ve done a lot more development work of my Pressworks control surface as a docked side panel in Muse. And I’ve done a lot of work around extracting “symbol units” from my corpus (think of them as maybe narrative primitives important to the canon), and now analyzing for and approving certain kinds of connections between units.

So within Muse lives my corpus itself, extracted promoted units and their relationships. All of that could be transferred to live in Codex/Work files instead, of course – but again, using Work to pull from the corpus (I assume) ends up being costly in usage. Hence the idea to split it all up.

Anyway, this ranting went on longer than I expected, but I had regular ChatGPT work up a description of my proposed orchestration plan, which I will use and see if it lives up to expectations around delegating for quality and to reduce usage.

Here is that text:

A proposed architecture is to split the book-production pipeline across three different AI environments, each handling the kind of work it is best suited for.

Muse acts as the orchestration and research layer. It holds the corpus, symbol units, promoted associations, project state, and placement logic. It can retrieve relevant material, build research packets, help generate premises, track approvals, and map approved images to manuscript passages.

Regular ChatGPT Plus chats handle the generation-heavy work: drafting the manuscript, revising text, generating interior images, and producing cover variations. These tasks use a separate usage pool from Work.

ChatGPT Work is reserved for the final production stage. It receives a mostly complete package containing the manuscript, approved images, placement instructions, and project metadata. Its job is then limited to DOCX assembly, text verification, rendering, visual QA, and export.

The resulting pipeline is:

Muse orchestrates and researches → ChatGPT generates text and images → Muse maps and validates assets → Work assembles and finishes

The main advantage is economic as much as technical. Long-running coordination and corpus work can stay in Muse, creative generation can use the regular ChatGPT Plus allowance, and the more constrained Work meter is only used for the small part of the process that actually benefits from agentic file handling and document production.

— GPT 5.6 Sol Medium