12 sessions on FFmpeg with a coding agent: the agent failed at least once in 10, and 2 finished without a human stepping in.
Takeaway: Next time I'd bake in a retry/resume path for the workers up front, since two of three codex workers died mid-task on SSE timeouts today — the thing that actually saved the round was that I'd told them to leave artifacts on disk, so I could salvage the partial forensics and grade data myself instead of restarting from zero.
Takeaway: Next time I'd put the full bokeh-fix-plus-breathing-room task in one prompt from the start instead of splitting it into scout/backup/implement stages — the orchestrator ended up doing that anyway.
Takeaway: Next time I'd put the taste-notes requirement up front instead of bolting it on after a full render cycle — the agent did solid disciplined work delegating to sub-workers and gating changes, but the real decisions were always going to need my eyes on the video first.
Takeaway: Next time I'd tell it upfront that melt and xmllint passing doesn't mean kdenlive itself will open the file cleanly — actually launching the real app was the check that mattered, not the automated gate.
Takeaway: Next time I'd have the handoff note call out the audio-leak risk explicitly so the worker doesn't need a full forensic pass to find it.
Takeaway: Next time I'd tell it upfront that our footage is mixed-resolution (4K plus native 1080x1920 phone clips) so the vertical bake uses relative crop math from the start instead of hardcoded 4K dimensions that broke 28 of 30 clips.
Takeaway: I'd bake the worker-timeout handling into the first prompt next time — Codex SSE timeouts killed 2 of 3 background workers today, and only salvaging their partial artifacts by eye saved the round.
Takeaway: I'd start the next session by pointing straight at the handoff file and the specific list of flagged clips instead of re-deriving triage state from scratch — that reconstruction step ate a lot of time.
Takeaway: Next time I'd put the cover-photo disposition check and the taste-notes review earlier in the same prompt, since the disposition bug (attached_pic=0) only got caught because I forced an ffprobe verification step right after muxing — that's the correction that actually mattered here.
Takeaway: Next time I'd tell the agent upfront that only the actual Kdenlive GUI validates file hashes and track playlist structure — melt and xmllint gave false confidence twice and I only caught the real bugs by opening the app myself and eyeballing the dialog.
Takeaway: I'd put the exact three taste notes and the agy chunking method in the first prompt so the orchestrator doesn't have to rediscover conventions from the handoff doc every round.
Takeaway: Next time I'd tell it upfront that source footage is mixed-resolution (not all 4K) so the vertical bake gets written resolution-independent from the start instead of two rounds of fixes.