AI Collapses Known-Shaped Work, Not Unknown-Shaped Work
A calibration heuristic for AI coding assistance: it collapses the cost of work whose form is already determined and barely dents work bottlenecked on feedback rather than typing — so on hard targets, build the debugging loop before generating code.
A useful heuristic for estimating where AI coding assistance actually changes a project's cost: it collapses the cost of *known-shaped* work and barely dents *unknown-shaped* work. Known-shaped work is work whose form is already determined and where mainly the typing remains — boilerplate, glue code, build system configuration, device tree entries, UI scaffolding, reading an unfamiliar codebase, and synthesising scattered tribal knowledge from forums and wikis into one coherent account. These are the tasks where a model is strongest, and a project made mostly of them can genuinely compress from months to weeks. Unknown-shaped work is bring-up debugging: timing races, DMA, cache coherency, a board that runs for forty seconds and then hangs. The bottleneck there is not typing, it is *feedback*. A blank screen with no serial output gives a model nothing to reason from, and generating more candidate code against no signal does not converge on anything. Those problems are solved by instrumentation and bisection, and the generation step is the cheap part either way. The corollary is practical: build the feedback loop before generating code. On an embedded target that means a serial console, network shell access, and a fast reflash cycle, established first. With a tight loop, model assistance compounds. Without one, the output is code nobody can evaluate — a failure mode adjacent to Capability Hallucination in LLM Agents and to the degradation described in Context Rot in Long AI Coding Sessions: Why Agents Get Worse as Context Fills. The heuristic also predicts where assistance will not help at all: constraints that are legal or physical rather than authorial. No amount of generation produces a licensed, panel-specific binary blob — see E-Ink Waveform Files: The NDA Blob Under Every Open E-Reader OS. Applied to e-reader work the split comes out clean: replacing a device's userland is largely known-shaped and a real multiplier, while a from-scratch kernel port mostly is not. See Replacing an E-Reader's Userland vs Writing an OS: Three Tiers.