362 sessions with Pi: the agent failed at least once in 120, and 193 finished without a human stepping in.
Takeaway: Next time I'd lead with the specific problem (input placeholder wrapping) rather than a broader 'sidebar needs help' framing, since the agent initially spent time on a height/empty-state fix I didn't actually need.
Takeaway: Next time I'd front-load the positioning (who this is for, what makes it different from a generic AI chatbot) before asking the agent to build anything, since the whole first landing page draft got mostly replaced once I explained the 'coding agent, but native to Excel' angle.
Takeaway: Next time I'd open by asking for a concrete UX audit with a prioritized list before jumping into changes, since the iterative 'please do' loop worked but cost a lot of exploration turns per round.
Takeaway: Next time I'd start by telling it to actually load the running page with agent-browser instead of trusting typecheck/build passing as proof the UI worked — that's what actually surfaced the real state and let it self-verify.
Takeaway: Next time I'd open with the constraint I ended up stating later: reuse the existing Files UI/tool, no separate assistant-only flow — that would've saved a full exploration-and-pitch round trip.
Takeaway: Next time I'd tell the agent up front to check I'm on the right branch before it commits, since the push-to-main step turned into an unplanned rebase/cherry-pick conflict resolution.
Takeaway: I'd probably ask for the triage output in a more structured format next time, like a table with priority rankings, since the freeform review was useful but a bit dense to parse.
Takeaway: Next time I'd ask for the worktree isolation and the full install/CORS rewrite in one go, since that's basically what happened and it went smoothly with the agent doing its own research before editing.
Takeaway: Next time I'd start the prompt by pointing straight at format-cells.ts and the multi-range support so the agent doesn't have to rediscover it through grep.
Takeaway: I'd start by asking it to triage the alerts by severity and propose fix options up front, since going back and forth on 'what would option A entail' for the SSRF fix cost time I could have saved by asking for a menu of mitigations in the first prompt.
Takeaway: Next time I'd ask upfront for the fix to include a manual verification step (actually reload and check tabs) before calling it done, since the first commit passed all checks but never actually wired the restore into the boot path.
Takeaway: Next time I'd start the prompt with the concrete symptom and my suspicion of the recent stream-proxy change, since that's exactly where the bug was.
Takeaway: I'd start next time by asking directly for a fetch()/network audit of tools and extensions before discussing CSP in the abstract - that grep for zero outbound fetch calls was what actually let me lock down connect-src confidently instead of guessing.
Takeaway: Next time I'd point straight at Pi's own compaction implementation as the reference from the start, since checking 'what does Pi do' twice (for token budgeting and for the temperature/reasoning conflict) was what actually resolved both the context % bug and the mysterious /compact 400 error.
Takeaway: Next time I'd give the agent my full list of ideas up front, including the file-storage and multi-workbook points I only added later, so it doesn't have to file issues in two separate batches.
Takeaway: Next time I'd open with both threads at once — the distribution manifest question and the view tool scope — since neither ended in code changes, just investigation and a filed issue.
Takeaway: Next time I'd bundle the alignment/width and pill-consistency asks into one clear spec upfront, since they took many small back-and-forth rounds of 'now this doesn't match that' across tool cards, thinking blocks and user messages.
Takeaway: Next time I'd ask upfront for a screenshot-driven audit at the actual sidebar width before any rebuild, since the real bug (an unlayered CSS reset clobbering Tailwind's @layer utilities) was hiding under a much bigger 'replace the whole layout shell' rewrite.
Takeaway: Next time I'd give the agent the correct repo path up front since it wasted several tool calls guessing wrong locations before finding the code.