Real Pi sessions
362 sessions with Pi: the agent failed at least once in 120, and 193 finished without a human stepping in.
Add inline setup card for web search failures in Excel add-in
Takeaway: Next time I'd front-load the dev-mode vs production proxy-probing distinction in the first prompt, since that's exactly the false-positive Codex caught later.
Investigate Excel TABLE() array formula support via Office.js API
Takeaway: I went in assuming there might be some Office.js trick to trigger TABLE(), but the honest answer was no — it's genuinely UI-only.
Allow extension slash commands to run while Pi is streaming
Takeaway: Next time I'd ask upfront for the busyAllowed default and mention this is pre-release so it can just default to true, since that saved a back-and-forth.
Fix currency text triggering KaTeX math in Excel add-in thinking blocks
Takeaway: Next time I'd start the prompt by pointing straight at markdown-block's KaTeX inline math extension instead of letting the agent rediscover it through a long grep trail — that diagnosis phase took way longer than the actual fix.
Fix responsive status footer layout in Excel add-in sidebar
Takeaway: Next time I'd start by describing the target layout (two lanes, mode badge always right) up front instead of iterating screenshot by screenshot — that would have saved a round trip.
Fix chat autoscroll jitter and inner scroll during streaming (Excel add-in)
Takeaway: Next time I'd ask for the holistic design pass (outer vs inner scroll UX) up front instead of shipping a quick distance-heuristic fix first — that first attempt (#401) looked fine in CI but didn't actually work in practice, and I only caught it by testing manually.
Design VBA execution support for Excel add-in agent
Takeaway: Next time I'd open with the real requirement up front: 'agent writes and runs arbitrary VBA like it does Python' instead of letting the conversation drift into macro-allowlist framing first.
Draft HN reply explaining Office.js add-in vs VM/Wine
Takeaway: I'd start by pointing the agent straight at the README and install docs since that's where the real architecture explanation lives, rather than letting it grep around first.
Fix Excel add-in footer buttons becoming unclickable when composer has text
Takeaway: I'd keep the prompt as-is since it worked well — the isolated worktree plus agent-browser reproduction step is what actually caught the real bug (the textarea's field-sizing overlapping the footer hitbox) rather than guessing.
Restore broken Files overlay actions and refactor detail-action code
Takeaway: I'd start the next prompt by explicitly scoping whether I want just the fix or the fix-plus-refactor loop in one go, since the refactor pass doubled the session length.
Install and scope pi extension go-to-bed from agent-stuff repo
Takeaway: Next time I'd ask upfront to install only the specific extension rather than the whole package, since the agent installed everything first and then had to go back and filter settings.json down to just go-to-bed.
Review and merge Excel add-in PR #357, fix failing tests
Takeaway: I'd start the prompt by pointing straight at the failing CI job so the agent doesn't have to dig through logs first.
Investigate Excel taskpane auto-resize and manifest format tradeoffs
Takeaway: I'd start by asking directly whether Office even supports dynamic taskpane resizing before digging into the code, since that's a platform limitation no amount of code archaeology would fix.
Fix mkcert -install error in pi-for-excel-proxy CLI
Takeaway: Next time I'd start by telling the agent to inspect the actual repo and reproduce the bug immediately, instead of letting it spend a few turns suggesting generic workarounds like 'use a different npx version' or 'run npm run dev instead'.
Map the Excel add-in's UX flow
Takeaway: I'd start by pointing the agent straight at the taskpane entrypoint and init files instead of letting it wander the whole src tree first — it got there anyway but took a lot of exploration turns.
Replace window.confirm with overlay dialogs in Excel add-in
Takeaway: Next time I'd front-load the exact list of window.confirm call sites in the first prompt so the agent doesn't have to rediscover them twice across two PRs.
Clarify workspace memory contract and add compaction nudge in Pi for Excel
Takeaway: Next time I'd start the first prompt with the full context I already had in mind (rules vs notes vs memory, cross-workbook scoping) so the agent skips the long codebase archaeology phase before it can even discuss the design.
Bump Pi SDK dependencies (pi-ai, pi-agent-core, pi-web-ui) to 0.52.12
Takeaway: I'd start the next prompt by giving it the exact gh pr create body as a file from the start, since inline multi-line strings with backticks in bash keep breaking the command.