In an early episode of Conlangery (#4, I think?) Bianca espouses a generative approach to conlanging rather than a typological satisfaction approach; she generates text that she likes the appearance and sound of, then attempts to decode it, making revisions until a balance between a semantically consistent architecture and an appropriately euphonic (and/or cacophonic) result is found. This approach gives you immediate control over the look and feel of the product, and is something I've found very tempting to employ. I plan on doing it for Paligu, rather than fabricating a table of affixes blindly in advance, and I think it will produce a much more natural-feeling result.

Obviously this approach has its limits and pitfalls—particularly, substantial compromises may be necessary to ensure historically plausible derivations from ancestral forms (and thereby avoid the great "negatives-are-the-same-word-spelled-backwards" problem and similar.) For the generation of a proto-language, however, I think this would be an excellent way to escape the database-filling rut—of which I have my own obsessions, which are mostly about filling up dictionaries. (But that being said, dictionaries are familiar, useful things that people use all the time; what in the world gives conlangers such an itch to always produce phonologies first?! Are linguistics courses to blame? Is it just a meme in the community? If neither, who the hell uses WALS that much? How much can we blame this on ASDs? etc.)

I think it might make a neat community project for here at Annie to try and write a guide for this... let's call it gestalt conlanging, maybe? Sort of an alternative to the LCK targeted at more artsy types and less at analysis fetishists. Helge Fauskanger's introduction to Quenya has a grammar section that goes on for pages and pages about the inflections before it introduces a list of them, so clearly there are some very prominent materials which endorse some kind of deviation from the tabular norm.