I've thought about it a bunch over the years and my advice is to follow a few principles: procedural elements should almost always impact how the player plays, procedural elements should cross-interact with each other as much as possible, and procedural elements should be understandable to the player.

As an example in No Man's Sky the player immediately starts with a jetpack. It makes it so the terrain is approached the same way every time: you just jetpack over it. Very rarely something is tall enough to require you to walk around.

Legend of Zelda: Breath of the Wild is a positive counter example. You have stamina so you're constantly thinking about the height of cliff faces, you can glide so you're thinking about height points and straight lines through the world, cliff faces cannot be climbed in rain so you think about the weather, slopes can be slid down on a shield, etc. There are a bunch of systems that may you think about the topology of the terrain, so even just varying the height map becomes interesting.

Similarly in No Man's Sky there are different resources but they work the same way on each planet and the player interacts with them pretty much the same way everywhere.

When I say elements should be "understandable" what I mean is that a trap, which Dwarf Fortress falls into, is to add a vast simulation which the player feels they can never really grasp. Simulations should not be a black box because then they begin to feel like a random number generator.

The aim shouldn't be to make an "infinite" game but to make a very interesting system for the player to slowly understand and master. The developer shouldn't just add stuff, but they should think about the long journey (and eventual mastery) a player might take as they spend significant time with the systems.

The advice I've heard is 1) have as many different dimensions of variation as you can manage, and try to have them matter, 2) resist the urge to impart correlations into the generation, since this just reduces variety, and 3) allow your parameters to become extreme occasionally, so that some parts of your generation are notably different, and special because they're rare.

1) is the really hard part, I think: it's a challenge to come up with enough meaningful variations that also work from a gameplay point of view.

Spelunky and Terraria do a good job with Procgen.

Spelunky uses handcrafted chunks that are connected together like a puzzle. Each generation feels interesting because each chunk is interesting and they are just strung together into a unique maze.

Terraria does a similar thing, where all the things that matter are hand-crafted. Bosses, items, layer details, progression. The map and cave system is generated but the caves tend to string together hand-crafted experiences and hand-crafted monsters. Many areas are generated, but in a unique style for variety. And there was still a fixed progression in the background (ore and boss progression). I think its in a big contrast to a similar game like Starbound, where all the weapons and creatures and maps were procgen at first, and it felt very '10000 bowls of oatmeal'. Terraria vs Starbound is the most like-for-like comparison I've seen for "procgen done right vs wrong"

Roguelikes are the ultimate in procgen technology, and should be the basis of any design. (DF is the ultimate simulation but uses procgen as a tool to achieve it; a different vein than NMS trying to explore the procgen and its direct outcomes itself)

1. Any imaginary world needs to be internally consistent to be interesting. Procgen does not void this rule; randomization needs to be gated to that internal consistency.

2. A world can be composed of multiple environments, and each environment can a different definition of internal consistency. Dungeon Crawl Stone Soup does this very openly and to fairly good effect, but pretty much any decent procgen game will inevitably end up with the same choice, because it's what enables you to play with various procgen systems/rules while still having a coherent overall game.

3. Generalizing 2, procgen systems generally do not start from nothing; You embed reasonably well-defined bases (races, weapon forms, etc) and then apply randomization on top of those bases to form the final element. Meaningful fundamental variation in those bases leads to guaranteed meaningful variation in the end result.

4. Games are fundamentally interactive systems. Generating a thousand variations without actually changing interaction is equivalent to generating exactly one variation. A blue orc vs a red orc is a not real variant, unless you make it so by attaching effects / statuses / capabilities. The color distinction is then a signal into the more abstracted changes that have occurred. This is where NMS falls over -- the images have changed, but the interaction has not.

5. A corollary to point 4, changing numbers without changing interaction is equivalent to doing nothing. If a meaningful damage threshold is functionally every 10 points of damage, then generating numbers 0-9, 11-19, etc is equivalent to generating just the numbers 0 and 10. This is where borderlands falls over -- a million variations in guns, but the vast majority of those variations are just numeric differences that don't change anything in practice.

6. If the procgen applies to equipment/builds, you pretty much need permadeath as a driver for exploring a procgen game and its systems, because you need players to explore for a local optima. Exploring the space for a global optima in a procgen game inevitably devolves into a gambling simulator, because players optimize out the fun, and the optimal play is to the kill unit XYZ repeatedly for the 0.5% drop. Games like Terraria fall into this trap (though Terraria resolves it mostly by making unit XYZ be trivially summon-able bosses)

7. Procgen can never guarantee interesting results; You inevitably need a system of vaults/artifacts -- hardcoded quality items -- to guarantee periodic interesting locations/items. Ideally more notable than average generation, but less than strictly optimal so as to not overpower the other systems -- an issue I've had with games like Diablo 3, where finding an item-set was vastly more powerful than anything else that I acquired for the next 5+ hours, eliminating the entire looting system (and even character skill system, since it forced my build too)

8. Decent Procgen is inevitably rule-based, with possibly multiple chained-rules each doing their own rolls, to ensure coherency. Ideally the rule-chain is roughly discoverable from the end results (by virtue of following internally consistent and apparent rules) -- an end-result that cannot be explained is indistinguishable from pure-random, and pure-random in incoherent. See any Tabletop RPG generation tables.

9. Hell, just start any procgen design from tabletop RPG tables.

10. The number of variations that can be produced is meaningless. This is only for marketing.

11. A procedural system that generates exactly once is fairly pointless. Permadeath.

12. A procedural system is inherently unfair, because you cannot guarantee difficulty, proper placement, unit composition, etc without removing the procedural systems outright. Embrace the unfairness, and design around it. Permadeath.

13. A procedural system is inherently unfair; you need to give some guaranteed items with guaranteed results to navigate an uncertain world (e.g. escape tools).

14. It's always good to enable some control by the player over the procedural generation. Choice of environments, chances at additional loot, etc. You want the player to engage with procgen rules, because they're interesting (and if they're not, what are you doing?), and the easiest way is to give them tools to manipulate it to their advantage. Also permadeath.

You want to look at dwarf fortress talks for that, I think?