AI Research Feedback
AI Research Feedback is a collection of multi-agent Claude Code skills that referee your economics or finance paper, pre-analysis plan, grant proposal, or paper-code alignment before you submit.
Last reviewed by the paperbanana team on Jul 13, 2026
Install
mkdir -p ~/.claude/skills/review-paper && curl -o ~/.claude/skills/review-paper/SKILL.md https://raw.githubusercontent.com/claesbackman/AI-research-feedback/main/Skills/review-paper/SKILL.mdWhat is AI Research Feedback?
AI Research Feedback is a collection of Claude Code skills for academic research review, developed by Claes Bäckman. Rather than one tool, it ships five separate, independently installable skills, each living in its own folder with its own SKILL.md: review-paper (full 8-agent referee-style review), review-paper-light (fast 2-agent check), review-paper-code (paper-code reproducibility and alignment review), review-pap (pre-analysis plan review), and review-grant (grant proposal review). Each skill sets disable-model-invocation: true, so a review only runs when you explicitly type its slash command — Claude will never launch a multi-agent review on its own.
The flagship review-paper skill runs eight specialized review agents in parallel, covering spelling/grammar/style, internal consistency and cross-reference verification, unsupported claims, mathematics and notation, tables and figures, referee assessment against a named journal's standards, and — distinctively — two agents that independently rate the paper's central contribution from opposite directions: a steelman advocate and a skeptic attacker, both grounded strictly in the paper's own bibliography. A synthesis section then reconciles the two, naming where they agree and the crux of any disagreement, and flags novelty claims the bibliography cannot verify. The skill supports named journal personas across top-5 economics (AER, QJE, JPE, Econometrica, REStud), finance (JF, JFE, RFS, JFQA), and macro (AEJMacro, JME, RED) journals.
The other four skills cover adjacent needs at different depths: review-paper-light gives a fast 2-agent pass on contribution, identification, and causal overclaiming for quick iteration; review-paper-code discovers your LaTeX paper and Stata/R/Python analysis code, checks reproducibility (paths, seeds, run order, dependencies), and maps empirical claims to code; review-pap runs a 6-agent pre-analysis plan review against trial registries (AEA, EGAP, OSF, ClinicalTrials, ISRCTN) or journal standards; and review-grant runs a 6-agent panel review of a grant proposal against funders like NSF, NIH, ERC, or HorizonEurope. All five require Claude Code with access to the general-purpose subagent.
Core capabilities
8-agent full referee review (review-paper)
Runs style, consistency, claims-integrity, math/notation, tables/figures, referee-assessment, and two opposed contribution-rating agents (advocate vs. skeptic) in parallel, then reconciles them into one consolidated report.
Journal-persona simulation
Recognizes named journals across top-5 economics, finance, and macro fields (e.g. /review-paper QJE, /review-paper JF) and applies that journal's editorial standards; falls back to high general standards if no journal is given.
Fast 2-agent quick check (review-paper-light)
A lighter, faster pass focused on contribution, identification, and causal overclaiming — meant for quick iteration before running the full 8-agent review.
Paper-code reproducibility review (review-paper-code)
Discovers the main .tex file and Stata/R/Python analysis scripts, checks reproducibility signals (paths, seeds, outputs, run order, documentation) and code quality, then maps tables, variables, and methods back to the paper text.
Pre-analysis plan review (review-pap)
6-agent review of a PAP's writing quality, specification completeness, internal consistency, identification strategy, statistical analysis, and registry/journal fit, with support for power calculations, survey instruments, and randomization protocols.
Grant proposal review (review-grant)
6-agent panel review evaluating clarity, significance, innovation, research design, feasibility, budget logic, and team readiness against a named funder or program, plus supporting materials like budgets and biosketches.
What you can use it for
Pre-submission referee report for a target journal
Run /review-paper QJE (or JF, AER, Econometrica, etc.) before submitting to get a consolidated referee report with contribution advocate/skeptic reconciliation, saved to reviews/PRE_SUBMISSION_REVIEW_[date].md.
Quick sanity check between drafts
Use /review-paper-light during iteration to catch causal overclaiming or unsupported identification claims fast, without running the full 8-agent pipeline every time.
Verifying an empirical paper matches its code
Run /review-paper-code path/to/main.tex path/to/code_dir full to check that tables, sample restrictions, and fixed-effects specifications in the paper actually match what the analysis scripts compute.
Reviewing a pre-analysis plan before registry submission
Run /review-pap AEA (or another registry/journal target) to get a 6-agent assessment of specification completeness and identification strategy before locking in a PAP.
Stress-testing a grant proposal before submission
Run /review-grant NSF or /review-grant NIH to get panel-style feedback on significance, feasibility, and budget logic ahead of a real submission deadline.
How to get started
- 1
Install the skills you need
For each skill, run mkdir -p ~/.claude/skills/<name> && curl -o ~/.claude/skills/<name>/SKILL.md https://raw.githubusercontent.com/claesbackman/AI-research-feedback/main/Skills/<name>/SKILL.md, substituting review-paper, review-paper-light, review-paper-code, review-pap, or review-grant.
- 2
Restart Claude Code so skills are picked up
Global installs under ~/.claude/skills/ are available in every project; project-local installs under .claude/skills/ apply only to that repo. Skills load the next time you start Claude Code in the target directory.
- 3
Invoke explicitly with the slash command
Type /review-paper, optionally with a journal code and/or path, e.g. /review-paper JF path/to/main.tex. Each skill has disable-model-invocation: true, so it never runs automatically.
- 4
Read the versioned report
Each skill saves its output into a reviews/ subfolder (e.g. reviews/PRE_SUBMISSION_REVIEW_[YYYY-MM-DD].md) and auto-versions the filename (-v2, -v3, ...) if a report for that date already exists, so past reviews never overwrite each other.
How it compares to similar skills
AI Research Feedback is a review-only tool set focused on economics, finance, and empirical social science — it does not draft papers for you. For drafting or broader-discipline needs, pair it with one of these.
Academic Research Skills
Pick Academic Research Skills if you need a full research → write → review → revise pipeline rather than a standalone review pass on an already-drafted paper.
PaperOrchestra
Pick PaperOrchestra if you need the actual drafting pipeline (outline, plotting, literature review, section writing) before you get to the review stage AI Research Feedback covers.
Chinese Thesis Workbench Skill
Pick Chinese Thesis Workbench Skill if your document is a Chinese-language undergraduate thesis rather than an English-language economics/finance journal submission.