I’d be a bit careful with the assumption that a good enough overarching doc stops the create feature getting rewritten when bulk-edit turns up. In my experience it usually doesn’t. Trying to design the whole thing up front so the later feature drops in cleanly is the old big-design-up-front problem, you end up guessing about the part you haven’t built. souperk and doo are right that it’s cheaper to expect the refactor and keep the steps small.

What the doc is genuinely useful for is the boundaries. Write down what the create part is allowed to depend on and hold it to that. If create only ever goes through a repo abstraction and never reaches into provisioning internals it doesn’t need, then bulk-edit comes in as a new module against the same boundary instead of a rewrite. When you do find the second feature forcing changes deep in the first, that’s nearly always the first one having coupled to something it never needed, and no amount of spec detail up front would have caught it. So I’d keep the design phase you’re after, just spend it on those boundaries rather than trying to predict the features you haven’t specced.

I wrote the boundaries idea up here if it’s any use: https://prickles.org/tenet/bounded-contexts/A6

There’s a useful split lurking in this. For narrow agentic work like retrieval over internal docs, structured classification, test scaffolding, deterministic refactor passes, a self-hosted 30B-class model can be fine and the inference economics work out at team scale. For multi-step planning and the harder agent loops, the frontier gap still shows up in the number of retries and the time-to-correct-answer.

The honest test is to pick the prompt category that’s costing you the most and benchmark something like Qwen 2.5 Coder 32B or DeepSeek V3 against whatever you’re paying for now. If the gap is small you’ve found your candidate. If it isn’t, you’ve at least costed the gap accurately rather than guessing at it.

The two costs people underestimate are the GPU box (plus a second one for the eval/staging path) and the maintenance overhead. Model picks go stale fast and someone on the team has to own that, or you end up shipping a Llama 3.1 stack into 2026 because nobody rebuilt the harness for whatever’s current.

The networking advice is right, but here’s the part you can actually control while you wait for that to pay off. Most fresh-grad applications look the same from the other side of the desk, a degree and a list of tutorials. What made me put someone through to interview was evidence of judgment: one or two small projects that are genuinely finished, with tests and a readme, ideally deployed somewhere I can poke at. Not a half-built clone of something. And in the cover note, one specific trade-off you made and why you made it. That reads as someone who has actually shipped, which is rarer than it should be.

I’ve used spec-kit and also rolled my own lighter version of the same idea. The framework matters less than the habit it forces, which is writing down what ‘done’ looks like before the agent starts. When I just prompt without that, I get something plausible that I then argue with for an hour. When the spec has acceptance criteria the agent can check itself against, most of that back-and-forth goes away. The caveat is that spec-kit can be heavier than a small change needs, so for quick work I’ll write the criteria in a comment and skip the ceremony. Worth trying. For me the habit of writing the criteria first is what stuck, more than the tooling around it.

squaresinger’s point matches what I’ve found. Once three agents are going, you become the coordination point - you’re holding the plan and reviewing all of it, and that part doesn’t scale the way the generating does. What’s kept it manageable for me is treating each one like an intern on a single, well-specified task I can check before it moves on, rather than running a swarm and hoping it converges. Wrote this up here: https://prickles.org/tenet/the-intern-pattern/AI1