Claude Code as a Grumpy, Peer-Reviewing Contractor

What it does

This is a configuration template and rule set for Anthropic’s Claude Code, tuned for academic work like lecture slides, research papers, and replication packages. You describe the project; Claude drafts a specification, invokes specialized skills, and runs the output through focused reviewer agents before presenting results. It is essentially a harness that turns a general coding assistant into a domain-specific workflow tool with memory, hooks, and quality thresholds.

The interesting bit

Instead of one generalist pass, the workflow spins up 18 narrow agents—proofreader, slide-auditor, pedagogy-reviewer, methods-referee—and makes them argue. An adversarial critic and fixer loop up to five rounds, and a Chain-of-Verification step forks a verifier that has never seen the draft to catch fabricated citations or numerical contradictions. The README is admirably honest that quality thresholds are advisory, not enforced by hooks; bypass the skill and you bypass the review.

Key highlights

• 18 specialized agents review distinct dimensions (grammar, pedagogy, R code, domain correctness) rather than relying on a single generalist.
• Adversarial QA pits a critic against a fixer until the critic approves or the round limit hits.
• Numeric scoring with advisory thresholds at 80, 90, and 95 for commit, PR, and excellence levels.
• Reproducibility auditing cross-checks every numeric claim in a manuscript against the script that produced it, tracked in a per-paper passport.yaml.
• Context survival via PreCompact hooks and a committed MEMORY.md that accumulates learnings across sessions.

Caveats

• Quality gates are harness-level only; a standard repository commit bypasses all checks, and the README notes you must configure hard enforcement yourself.
• The template is heavily optimized for LaTeX/Beamer and Quarto; while the orchestration patterns work for any text artifact, the out-of-the-box skills and CLAUDE.md assume an academic toolchain.

Verdict

Academics and researchers already using Claude Code who want structured, repeatable quality control for manuscripts and slides will find this a rare example of workflow rigor. If you are not writing LaTeX or managing replication packages, much of the included scaffolding will be irrelevant.