aris-auto-paper-improvement-loop

Run this skill on a compiled paper directory containing main.pdf and LaTeX sources. It sends the full text to GPT-5.4 for a structured review, implements fixes by severity, recompiles with latexmk, and repeats for two rounds.

It runs exactly two rounds of review, fix, and recompile. Testing shows Round 1 catches structural issues and Round 2 catches remaining presentation issues, with diminishing returns beyond two rounds for writing-only improvements.

Yes, enable the human checkpoint option to pause after each round's review. You can approve all fixes, give custom modification instructions, skip specific fixes, or stop the loop early.

The skill writes PAPER_IMPROVEMENT_STATE.json after each round with the current round, thread ID, and score. On restart, it reads this file plus the improvement log to resume from the next round if the state is less than 24 hours old.

Yes, after the final recompilation it checks page count against venue limits and scans the log for overfull and underfull hbox warnings. It applies fixes such as \resizebox, font reduction, or rephrasing, then recompiles.

No, the skill explicitly forbids fabricating experimental results. It softens overclaims, fixes inconsistencies, and may add synthetic validation subsections describing methodology, but never invents unsupported numbers or claims.