Building with AI on a slow connection

Most AI advice assumes a fast, stable connection. Building from Nsukka, I've had to design my entire workflow around the reality that the network will fail me at the worst moment.

The habits that keep me shipping

• batch work before going online: collect all my questions, then make one dense session out of them, instead of a hundred tiny round-trips
• keep prompts and answers in plain files — a local log of every exchange, so a dropped request never loses thinking work
• prefer tools with retry and resume; anything that punishes a flaky connection gets deleted from my workflow fast
• download the heavy things once: model docs, style guides, reference repos — the network is for interaction, not for re-reading

The unexpected upside

Bad networks force the discipline good engineers recommend anyway: write the task down clearly before you send it, keep local copies of everything, design features that degrade gracefully. The outage is the test. If my day collapses without connectivity, my workflow was fragile, not my connection.

There's a deeper point here for anyone building AI products for Africa: assume intermittency. Cache aggressively, queue writes, make the offline path a feature instead of an apology. That constraint produces better software, not worse.