Always re-run from the original, never from the last attempt
The instinct when a result is still too large is to compress that result again. Doing so stacks generation loss: the second pass re-encodes images that already carry the first pass artefacts, so the artefacts are baked in and then approximated again. Two moderate passes look considerably worse than one firmer pass at the same final size.
Keep the original file in place and change the settings instead. Each attempt is then a single generation from clean source data, and you can compare candidates honestly because they differ only in settings, not in how many times they have been through the encoder.