PrettySeriousResearcher: Skill for Claude Code

Research mock-review — the pre-submission reviewer gate

Simulate the review the paper is about to face, so the agent surfaces the weaknesses a real reviewer would raise before submission — and hands the user a ranked, pre-critiqued decision instead of an open-ended "please review this." This is the single biggest lever for reducing the user's review burden: the diagnosis is done for them; they confirm and choose.

The autonomy inversion (read first)

The strong autonomous-science systems (AI-Scientist's ensemble reviewer, Agent Laboratory's reviewer, Zochi, Agents4Science's LLM panel) prove the machinery works — but most of them auto-decide accept/reject. We invert that: the panel scores, ranks, and drafts; the user stays the chooser. This skill never emits an accept/reject verdict, never gates submission by itself, and never edits the paper — it produces a packet the user acts on. Auto-deciding is the opposite of the design philosophy.

When it fires

• Phase-4 exit (primary): a full pass on the finished draft, as part of the exit criteria.
• Phase-1 dry-run (lightweight): a single-reviewer pass on the skeleton to fix the target bar early — what score must this venue's rubric reach, and which dimensions matter most — so every later phase optimizes toward it. Uses the venue profile from research-venue-selection.
• On request / before any resubmission.

The panel: driven via coding-agent CLIs, not API keys

The panel is realized by driving installed coding-agent CLIs/SDKs as subprocess reviewers — e.g. claude -p "<review prompt>", codex exec "<review prompt>", gemini -p "<review prompt>" — each producing an independent review captured from stdout. This is deliberate:

• No API keys, no secrets. Each CLI owns its own auth; this skill never asks for, stores, or logs a key (keeps the install_skills.py secrets-lint clean and the set portable).
• Detect billing before launch. A CLI may be subscription-covered, credit-limited, metered, or API-backed. Confirm the basis for every panel member; book subscription-covered use to its plan and explicit metered charges as coding-agent actuals. Keep panel runtime separate from experiment spend.
• Reviewer family != writer family. Assign each review to a different family than the one that wrote the draft to kill the known self-preference / sycophancy bias. Detect which CLIs are on PATH; use the diverse set.
• Graceful degradation. If only one family is available, run multiple personas/roles at varied temperature instead — but then treat agreement with suspicion and discount the panel's confidence (a same-family panel is an echo chamber, not independent evidence).