Автономная доработка академической статьи

★ 7.1 · research

auto-paper-improvement-loop is a Claude Code skill that autonomously improves a compiled academic paper through two rounds of external LLM review, fix implementation, and recompilation. Round 1 addresses structural issues (score 4→6/10), Round 2 catches remaining presentation problems (6→7/10), with diminishing returns beyond that. Each review round opens a fresh Codex MCP thread via GPT-5.6-Sol — a deliberate bias guard, since reusing the same thread with prior context was observed to inflate scores from a true 3/10 to a false 8/10. The skill accepts a paper directory, an optional `--style-ref` source (local TeX, PDF, or arXiv ID) for structural nudges, and an optional `--edit-whitelist` path to constrain which files may be modified; all rounds and rejected edits are logged to `PAPER_IMPROVEMENT_LOG.md`. It is intended for researchers who want to automatically polish a LaTeX paper to publication quality after the initial compilation step.