A collection of Claude Code skills for reviewing and understanding academic research. This tool was developed by Claes Bäckman.
Each skill lives in its own folder containing a SKILL.md file:
Skills/review-paper/SKILL.md: Full 8-agent referee-style paper review.Skills/review-paper-light/SKILL.md: Fast 2-agent paper check.Skills/review-paper-code/SKILL.md: Paper–code reproducibility and alignment review.Skills/review-pap/SKILL.md: Pre-analysis plan review.Skills/review-grant/SKILL.md: Grant proposal review.Skills/explain-diff/SKILL.md: Explain a code change as an offline HTML page with a quiz.
These are Claude Code skills. Install one into ~/.claude/skills/<name>/SKILL.md and invoke it by typing /<name> (for example, /review-paper) inside Claude Code. The per-skill installation commands below create the required folder and download the SKILL.md for you.
A global install (~/.claude/skills/) is available in every project. A project-local install (.claude/skills/) applies only to that repository. Skills are picked up the next time you start Claude Code in the target directory.
Each skill sets disable-model-invocation: true, so a review runs only when you explicitly type its / command. Claude will never launch a multi-agent review on its own.
Already installed these as slash commands? Custom commands and skills have merged in Claude Code, so any existing
~/.claude/commands/<name>.mdfile keeps working and still provides/<name>. To avoid two definitions of the same command, delete the old~/.claude/commands/<name>.mdfile after installing the skill version.
Runs a rigorous pre-submission review of an academic paper, simulating the scrutiny of a specific journal's editorial board. Eight specialized review agents run in parallel and consolidate their findings into a single structured report. The review only reads the paper's own source files (the main .tex file and everything it includes) and ignores old drafts, response letters, and previous review reports in the same folder.
What it reviews:
| Agent | Focus |
|---|---|
| 1 | Spelling, grammar, and academic style |
| 2 | Internal consistency and cross-reference verification |
| 3 | Unsupported claims and identification integrity |
| 4 | Mathematics, equations, and notation |
| 5 | Tables, figures, and their documentation |
| 6 | Referee assessment (identification, analyses, positioning, journal fit) |
| 7 | Contribution advocate (steelman, grounded in the paper's own bibliography) |
| 8 | Contribution skeptic (attack, grounded in the paper's own bibliography) |
Agents 7 and 8 independently rate the central contribution from opposite directions. The report reconciles them in a synthesis section that states where they agree, names the crux of any disagreement, and flags novelty claims that cannot be verified from the paper's own bibliography.
Installation:
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.mdFor a project-local 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.mdUsage:
/review-paper
/review-paper QJE
/review-paper JF path/to/main.tex
Supported journals:
| Category | Journals |
|---|---|
| Top-5 economics | AER, QJE, JPE, Econometrica, REStud |
| Finance | JF, JFE, RFS, JFQA |
| Macro | AEJMacro, JME, RED |
If no journal is specified, the skill applies high general standards without a specific journal persona. If no path is provided, it auto-detects the main .tex file.
Output:
Saves a consolidated report to reviews/PRE_SUBMISSION_REVIEW_[YYYY-MM-DD].md, automatically appending -v2, -v3, and so on if a file already exists. Reports live in a reviews/ subfolder so they cannot contaminate future runs.
Customization:
- Add journals or fields by editing the recognized journal names list in the skill file.
- Add project-specific context in your prompt or in a local
CLAUDE.mdfile. - Adjust folder discovery or save paths directly in the skill if your project structure differs from the default assumptions.
Requirements:
- Claude Code with access to the
general-purposesubagent. - A LaTeX paper. The skill reads
.texfiles and optionally inspects figure and table files.
Runs a fast 2-agent pre-submission check for an economics paper. It focuses on contribution, identification, causal overclaiming, and unsupported claims, and is designed for quick iteration before a full review.
Installation:
mkdir -p ~/.claude/skills/review-paper-light && curl -o ~/.claude/skills/review-paper-light/SKILL.md \
https://raw.githubusercontent.com/claesbackman/AI-research-feedback/main/Skills/review-paper-light/SKILL.mdFor a project-local install:
mkdir -p .claude/skills/review-paper-light && curl -o .claude/skills/review-paper-light/SKILL.md \
https://raw.githubusercontent.com/claesbackman/AI-research-feedback/main/Skills/review-paper-light/SKILL.mdUsage:
/review-paper-light
/review-paper-light path/to/main.tex
If no path is provided, the skill auto-detects the main .tex file.
Output:
Saves a short prioritized report to reviews/QUICK_REVIEW_[YYYY-MM-DD].md, automatically versioning the filename if one already exists.
Requirements:
- Claude Code with access to the
general-purposesubagent. - A LaTeX paper.
Runs a paper-code review for empirical research projects. It discovers the main LaTeX paper and analysis code, checks reproducibility and code quality, maps the paper's main empirical claims to the code, and writes a constructive report highlighting strengths, gaps to verify, and concrete next steps.
What it reviews:
| Area | Focus |
|---|---|
| Paper discovery | Main .tex file and included sections |
| Code discovery | Stata, R, and Python scripts in common analysis folders |
| Reproducibility | Paths, seeds, outputs, dependencies, run order, documentation |
| Code quality | Structure, commented-out code, opaque transforms, major thresholds |
| Paper-code alignment | Tables, variables, sample restrictions, methods, clustering, fixed effects |
Installation:
mkdir -p ~/.claude/skills/review-paper-code && curl -o ~/.claude/skills/review-paper-code/SKILL.md \
https://raw.githubusercontent.com/claesbackman/AI-research-feedback/main/Skills/review-paper-code/SKILL.mdFor a project-local install:
mkdir -p .claude/skills/review-paper-code && curl -o .claude/skills/review-paper-code/SKILL.md \
https://raw.githubusercontent.com/claesbackman/AI-research-feedback/main/Skills/review-paper-code/SKILL.mdUsage:
/review-paper-code
/review-paper-code path/to/main.tex
/review-paper-code path/to/main.tex path/to/code_dir
/review-paper-code path/to/main.tex path/to/code_dir full
Review depth:
main: default; focuses on main scripts and core outputsfull: reviews all detected code files in scope
Output:
Writes a report to reviews/code_review_report.md, creating the reviews/ folder if needed.
Requirements:
- Claude Code with access to the
general-purposesubagent. - A LaTeX paper plus Stata, R, or Python analysis code.
Runs a 6-agent pre-submission review of a pre-analysis plan (PAP). The skill auto-detects the main PAP and supporting files, then evaluates writing quality, specification completeness, internal consistency, identification strategy, statistical analysis, implementation details, and registry or journal fit.
Installation:
mkdir -p ~/.claude/skills/review-pap && curl -o ~/.claude/skills/review-pap/SKILL.md \
https://raw.githubusercontent.com/claesbackman/AI-research-feedback/main/Skills/review-pap/SKILL.mdFor a project-local install:
mkdir -p .claude/skills/review-pap && curl -o .claude/skills/review-pap/SKILL.md \
https://raw.githubusercontent.com/claesbackman/AI-research-feedback/main/Skills/review-pap/SKILL.mdUsage:
/review-pap
/review-pap AEA
/review-pap QJE path/to/pap.tex
Supported targets:
- Trial registries:
AEA,EGAP,OSF,ClinicalTrials,ISRCTN - Journal standards:
AER,QJE,JPE,RESTUD,AEJ,JEEA - General standards:
top-journal,working-paper
If no target is specified, the skill defaults to top-journal. If no path is provided, it auto-detects the main PAP file.
Supporting files it can inspect:
- Power calculations and sample-size worksheets
- Survey instruments and questionnaires
- Randomization protocols and sampling frames
- Code skeletons and mock tables
- Data dictionaries and ethics materials
Output:
Saves a consolidated report to reviews/PAP_REVIEW_[YYYY-MM-DD].md, automatically versioning the filename if one already exists.
Requirements:
- Claude Code with access to the
general-purposesubagent. - A PAP in a readable format such as
.md,.txt, or.tex. The skill can also attempt to work with.pdfand.docx, while noting accessibility limitations if needed.
Runs a 6-agent pre-submission panel review of a grant proposal. The skill auto-detects the main proposal and supporting documents, then evaluates clarity, compliance signals, internal consistency, significance, innovation, research design, feasibility, budget logic, team readiness, and fit to the target funder or program.
Installation:
mkdir -p ~/.claude/skills/review-grant && curl -o ~/.claude/skills/review-grant/SKILL.md \
https://raw.githubusercontent.com/claesbackman/AI-research-feedback/main/Skills/review-grant/SKILL.mdFor a project-local install:
mkdir -p .claude/skills/review-grant && curl -o .claude/skills/review-grant/SKILL.md \
https://raw.githubusercontent.com/claesbackman/AI-research-feedback/main/Skills/review-grant/SKILL.mdUsage:
/review-grant
/review-grant NSF
/review-grant NIH path/to/proposal.pdf
Supported funders/programs:
- US federal science and health:
NSF,NIH - International research funders:
ERC,HorizonEurope - General proposal standards:
major-funder,foundation
If no target is specified, the skill defaults to major-funder. If no path is provided, it auto-detects the main proposal file.
Supporting files it can inspect:
- Budgets and budget justifications
- Timelines and workplans
- Biosketches, CVs, and personnel documents
- Data-management plans, mentoring plans, and facilities statements
- Letters of support, appendices, and supplementary materials
Output:
Saves a consolidated report to reviews/GRANT_PROPOSAL_REVIEW_[YYYY-MM-DD].md, automatically versioning the filename if one already exists.
Requirements:
- Claude Code with access to the
general-purposesubagent. - A proposal in a readable format such as
.md,.txt, or.tex. The skill can also attempt to work with.pdfand.docx, while noting accessibility limitations if needed.
Explains a change to your analysis code well enough that someone who did not write it could defend it. This is the one skill here that is not a review. It is aimed at the coauthor who needs to understand what an AI assistant or a collaborator just did to the estimation pipeline, and at your future self returning to a project after six months.
The skill reads the diff and the surrounding code from scratch, ignoring any account of the change that already exists in the conversation, then writes a self-contained HTML page you open in a browser.
The idea comes from Geoffrey Litt's Understanding is the new bottleneck, which argues that once agents write most of the code, the scarce resource is not verification but the researcher's own grasp of the system — you need the concepts in your head to think about what to do next. Litt proposes explainer documents and embedded quizzes as a deliberate speed regulator. This skill is that proposal applied to empirical research code, where the thing you must not lose track of is what happened to the sample and the coefficients.
What the page contains:
| Section | Content |
|---|---|
| 1 | What the code did before, established by reading the surrounding scripts |
| 2 | What changed and why, in plain language with no code |
| 3 | Consequences for the results — which sample, which coefficients, which tables, and in which direction |
| 4 | Walkthrough of the changed code, grouped by purpose rather than by file |
| 5 | A five-question multiple-choice quiz with explanations for every option |
Section 3 is the point of the skill. It requires verification rather than inference: comparing observation counts and coefficients between runs, checking whether a modified table file differs in its numbers or only in a timestamp, and reading any new output file before quoting a figure from it. When nothing about the results changed, the page has to say so explicitly and show the evidence.
At least two quiz questions must test empirical consequences rather than syntax — which observations enter the sample, what the coefficient now identifies, what would change if an assumption failed.
Installation:
mkdir -p ~/.claude/skills/explain-diff && curl -o ~/.claude/skills/explain-diff/SKILL.md \
https://raw.githubusercontent.com/claesbackman/AI-research-feedback/main/Skills/explain-diff/SKILL.mdFor a project-local install:
mkdir -p .claude/skills/explain-diff && curl -o .claude/skills/explain-diff/SKILL.md \
https://raw.githubusercontent.com/claesbackman/AI-research-feedback/main/Skills/explain-diff/SKILL.mdUsage:
/explain-diff
/explain-diff abc123
/explain-diff main..HEAD
With no argument the skill explains the uncommitted working tree against HEAD, including files that are new and therefore invisible to git diff. A single ref is compared against the working tree. A range is used as given.
Output:
Writes one HTML file outside the repository, saved to ~/Documents/ unless you name another directory. Keeping the page out of the repo means it never lands in a commit or in your paper's folder, and the skill confirms with git status afterwards that nothing in the project changed. You also get a plain-text summary in the terminal, so you do not have to open the page to learn whether the results moved.
Filenames end in a ref tag so the pages stay matchable to commits as they accumulate: 2026-08-01-cluster-by-municipality-wt-e956a7e.html for a working tree against HEAD, or the short SHAs for a ref or a range. The page itself carries the full 40-character SHAs, the branch, and the exact comparison. When the comparison involves an uncommitted working tree, the page says so and lists the files, because that state is not reproducible later from the SHA alone.
Requirements:
- Claude Code. No subagents are used, so this skill runs in a single context.
- A git repository. Everything else — Stata, R, Python, LaTeX — is optional.
MIT — free to use, adapt, and share.