|
It is hard. Much harder than regular software work. A few things need to line up: model iteration speed, task chunk size, and your own context window and comprehension abilities. Too slow and you lose interest. Too easy and it's just tedious riffing. Too large and you're burned out by reviewing giant complicated walls of text. Etc. The bites have to be the right size and speed for both you and the LLM. In the few instances I've been able to achieve really joyful flow state, there are usually two simultaneous workstreams, plus or minus one. They're usually working towards a large goal that I roughly know how to judge, in digestible bites. For example, sequentially modifying the UI in a series of operations towards an overarching goal, where it's easy to tell if a step worked, and what the next step should be, but where you're not sure exactly what you want, so there's some curiosity and discovery rather than just feeding the bot tiny instructions. I try to keep the two workstreams from overlapping. If both streams start fighting over a file they both dirty, things go south fast. Adjusting your prefs/harness/etc for model terseness goes a LONG way. Context quality is absolutely everything. A good context "seed" can go back and forth for many turns cleanly and with focus, and even compact successfully once or twice. A bad seed will be annoying to work with from the jump, will thrash towards compaction faster, at which point it gets even worse. This is difficult to troubleshoot objectively, but I'll frequently restart conversations if I don't like the vibe of the first couple turns. It's made harder by constant model churn. Until opus 4.8, I ran opus/sonnet 4.5 high for a long time in large part for continuity of intuition, if that makes sense. There are also many elements of human knowledge management that make a difference. I've found "append only" to be a magic word, generating markdown logs of changes, or learnings, etc. Whatever workflows create visibility and resumability so that you can return from a spell away and get up to speed effectively. Manually keeping your own dev log alongside the session sometimes helps, makes things sticky and ensures you understand what's happening. But it's hard. Feels incredible when it goes well, but going well feels very nearly random, and whittling towards reproducibility can be very mentally draining, in terms of both energy and morale. |