Scaffolding Rots

Third blog post after 2 months of silence, already forgot how one expresses a fuzzy idea in words, but let’s give this a try.

I have been reflecting on how much tooling we have built on top of models to make them do what we want, but at the same time I have noticed less and less need for this tooling the smarter the models get: rules become unnecessary noise, hooks are redundant, and verification steps are baked in, and from the look of it, it will only keep on getting truer (looking at you fable, whenever you’re out).

The fear of lack of determinism in models has been pushing everyone who is heavily adopting them (including me) to build boundaries that promise the transformation of “manifestations from prompt” into “predictable outcomes”. For example, the Superpowers plugin had great success when it came out since it can help you brainstorm ideas (sometimes even visually in their latest versions) then uses proper PRD standards to write specs, then plans, then implementation with subagents. This, for me, “quickly” got replaced with a proper PROGRESS.md/STATE.md setup with model orchestration as soon as Opus 4.7 came out. The Ralph loop was promising people true agentic loops, but I prefer the idiomatic “/goal” in Claude with a proper verification process. Now that I’m thinking about it, I don’t even remember the last time I explicitly entered plan mode, and I like it that way.

I admit even for my usage they might not have gone to zero completely, and I can go for these in some cases, but they have definitely not survived the savior hype they promised.

The value from the scaffolding is shrinking, but the total volume of engineering is definitely exploding, aka Jevons’ paradox (technological improvements that increase the efficiency of a resource’s use lead to a rise, rather than a fall, in total consumption of that resource). And just like in the coal/steam version, because the steam engines used coal more efficiently, the usage of engines was more frequent, not the failure of the industries as speculated before.

The field grows, but volume going up doesn’t mean your particular skills are the ones in demand.

One thing I’ve actually found useful is learning how each model behaves.

Each model almost has a different personality in ways of tackling tech tasks and adapting to workflows. And sometimes you have to find the balance between creating a consistent or fit-for-all workflow and adapting to each model’s capabilities.