Thanks, Neil. Greetings after all these years!

In general, though, Fellou is a work in progress. It definitely needs better reporting on when/why a complex job fails, & on what deliverables are recoverable. And it's really frustrating that it doesn't return credits when a job fails. I ended up opting into a paid plan to get complex jobs done.

During a job, Fellou.ai deposits interim or in-progress deliverables in the following path on a Windows machine. You might look there for what's left behind after a failed job. C:\Users\[user_name] \AppData\Roaming\Fellou\FellouUserTempFileData

Also particularly impressive about asking ChatGPT to revise my prompts for Fellou is that I ended up with prompts that required results to be returned with higher data quality (in a taxonomy & data schema formatted in JSON; deduplicated; names standardized on canonical names, etc.)

Structuring a complex agentic research task in stages has two advantages. It sequences logical operations based on dependencies (e.g., what an agentic AI needs to do first to ensure reliability of next tasks); and it sets what are essentially debuggin breakpoints so I know about where a job fails.

Asking another LLM to optimize prompts produced (quoting ChatGPT) "tightened, agent-ready" prompts asking Fellou to "work in staged passes, enforce data quality, & deliver reusable assets" with "success criteria, schemas, checklists ... so it is far likelier to finish with high-quality outputs."

I've had success making complex Fellou.ai jobs run more reliably by uploading draft prompts to ChatGPT & asking for revisions optimized for operationalization ("optimize prompt to structure & stage Fellou's research & generation of deliverables with greatest chance of success & with best results").

Fellou cited 1 article for which I was lead author but collected many items from it, in the same way it collected 14 items from 1 of Geoffrey Harpham's pieces. My post discusses this at one point (see screenshot). Re: international balance, adjusting the prompt could change that, but I haven't tried

Hi, Elisa! Yes, both literals (which would seem to be context-free) & hierarchical or other relationships (requiring context to discern) seem to be problems for some AI use cases.

Here's the table of contents of The Humanities Laboratory: From Design to Archaeology. I recommend the editors’ introduction, “What Goes on in Humanities Laboratories?” (9-62). It’s truly wonderful: rich, deep, thoughtful.

“Liberal Arts Pantocrator” works into my argument short discussions of two early pictorial illustrations (by Herrad of Landsberg & Botticelli ) of the “seven liberal arts.”