
Takes a compiled LaTeX paper and runs two rounds of GPT-5.5 review, fix implementation, and recompilation without human intervention. Each round uses a fresh reviewer thread to avoid score inflation from context carryover, a real problem the author documented when trying reply-based reviews. You can optionally pass a style reference to guide structural fixes or an edit whitelist to lock down which files and operations are allowed, useful for resubmissions where you can't touch the bibliography or add new theorems. Set human checkpoint mode if you want to approve changes between rounds. The two round limit reflects diminishing returns: round one catches structural issues, round two catches presentation problems, and round three rarely improves the actual quality score.
$ npx -y skills add wanshuiyin/auto-claude-code-research-in-sleep --skill auto-paper-improvement-loop --agent claude-codeInstalls into .claude/skills of the current project.
Before each reviewer slot, follow Selectable Reviewers: resolve invocation →
.aris/reviewer.json→ Claude Code default (Codex MCP) withreviewer_route.pythrough the canonical helper resolver, preserving this slot's tier, fresh contexts, source/image inputs and deterministic gates. Codex examples/pins below apply only tobackend=codex; other routes use their own first-call/reply contract. Save the route and actual returned identity; unavailable selections areREVIEW_UNAVAILABLE. Classify acceptance from actual families/completion/scope. Forward invocation reviewer/model/effort overrides to reviewer-bearing children. Writer selection is separate; follow the writing-worker guidance below when drafting or revising prose.
Sponsors