Real sessions where developers added a feature to an existing codebase with a coding agent: the first prompt, the plan, where the agent drifted, and what shipped.
Things to look for in these sessions:
Takeaway: Next time I'd ask upfront how this OS's other interactive JS apps handle input, since the busy-poll getKeyNonBlock() loop turned out to be the real cause of the exit crash, not the VGA I/O port theory the agent chased first.
Takeaway: I'd tell it upfront that the tournament has already started and to verify current results before predicting, since it wasted several turns discovering that on its own via web searches.
Takeaway: Next time I'd give the exact repo URL and confirm gh auth up front — that cost us several minutes of dead-end guessing at the start.
Takeaway: Next time I'd tell it upfront that the official docs/install links might be stale or gated, since chasing the broken GitHub skill repo and rebuilding the skill from scratch ate a huge chunk of the session.
Takeaway: Next time I'd describe the corner/divider interaction more precisely upfront — screenshots after each change helped catch when the agent's approach (adding radius to the whole sidebar) visually broke things.
Takeaway: Next time I'd ask upfront for click-outside-to-close behavior along with the lightbox feature, since it's an obvious pairing.
Takeaway: Next time I'd describe the exact final look I want (no border, transparent by default, hover background) in one prompt instead of iterating three times.
Takeaway: Next time I'd be more specific upfront about exactly which padding/corners to change and preview before committing, since I ended up reverting two of the four style tweaks.
Takeaway: Next time I'd specify upfront whether I mean horizontal or vertical centering, since my first ask about 'centering' the title was ambiguous and the agent guessed wrong direction.
Takeaway: I'd specify next time whether the branch icon and spinner should stay, since that ambiguity took a few extra verification steps.
Takeaway: Next time I'd include a screenshot of the desired banner layout up front, since the first pass removed the right text but left an alignment issue I had to flag separately.
Takeaway: This was a one-line constant change, so I could have just asked for the exact constant name if I'd known it, but letting the agent grep for it first was fine too.
Takeaway: Next time I'd upfront ask for the facts/vector split explicitly instead of letting the agent build the monolithic GraphModality first and then retrofitting — that second prompt ('还有 factsworker和vectorworker没实现') is what actually forced the right architecture.
Takeaway: Next time I'd point directly at toolBlock.ts and ChatTranscript.vue upfront since I already knew roughly where tool rendering logic lived, instead of letting the agent grep broadly first.
Takeaway: Next time I'd open with the concrete ask up front — 'add a way to open any folder as a workspace, not just ones already scanned in .pi/agent/sessions' — instead of first describing the limitation and only later asking for the fix, since that cost a whole extra round of re-exploration.
Takeaway: Next time I'd specify upfront exactly which display elements (bar, token count, percentage, denominator) should and shouldn't carry the warning colour, since I ended up correcting that piecemeal over four separate prompts.
Takeaway: Next time I'd ask upfront for both the research and a concrete GitHub issue in one go, since the agent handled the investigation (codebase read, web research on Blargg's nes_ntsc and TetaNES) well without any correction needed from me.
Takeaway: I'd start by asking for the investigation and the two features (row dividers, longest-word minimum width) together, since that's basically what I ended up doing anyway.
Takeaway: I'd start by pointing straight at the powerline-footer extension path instead of letting the agent grep the whole monorepo for 'footer' and 'powerline' first — that exploration ate most of the session.