Software
- applied-micro-skills Claude Code and Codex
A collection of eighteen Claude Code and Codex skills for the full empirical applied-microeconomics research pipeline. The repository follows the way a paper actually develops: finding literature and data, constructing defensible variables, estimating models, producing publication-ready outputs, delivering manuscripts, and independently checking that the finished results follow from the underlying analysis. These are self-contained workflows, methodological guardrails, runnable tools, and documented fixes for errors that tend to surface only in serious research projects. Researchers can install the complete collection or select individual skills with no cross-skill dependencies.
- Literature, sources, and data. lit-review-orchestrator reads a manuscript, abstract, or proposal, derives a search plan, combines multiple deep-research and scholarly-search channels, deduplicates the results, and verifies that every retained paper exists before producing a ranked bibliography. wrds provides guarded access to Wharton Research Data Services, including fail-fast authentication that avoids locking an account through automatic login retries and practical query patterns for Compustat, CRSP, FactSet, and I/B/E/S. download-gated-pdfs handles sources that appear to be PDFs in a browser but return an HTML interstitial to automated clients, using archival routes to recover the underlying document when available.
- Defensible LLM-assisted classification. Five skills address methodological and operational risks in large labeling projects. annotator-input-parity-check asks whether a model receives the same evidence that human coders saw, because missing inputs impose a recall ceiling that prompt engineering cannot overcome. llm-gold-bound-failure-check tests whether an apparent model failure is actually created by noisy or inconsistent gold labels before another paid pilot is run. llm-campaign-drift-gate uses a fresh canary to detect provider-side changes during a paused campaign. adjudication-sheets reconstructs the complete, byte-identical evidence shown to the classifier so reviewers do not rule from truncated excerpts. asyncopenai-concurrency-httpx-pool removes the hidden connection-pool bottleneck that can cap supposedly high-concurrency batch scorers.
- Econometrics and computation. event-study-cars is a runnable Python engine for cumulative abnormal returns, validated against Stata's eventstudy2 across multiple specifications and equipped with thin-trading correction, market and factor models, boundary guards, and standard parametric and nonparametric tests. stata drives a licensed Stata installation through pystata and exchanges data with pandas in memory, making Stata analysis usable inside a broader Python workflow. pyfixest-cupy64-absorbed-regressors documents the coefficient-table and precision issues that arise when fixed-effects demeaning moves to the GPU, while pyfixest-grid-sharding shows how to divide large regression grids across workers when estimation remains CPU-bound despite GPU acceleration.
- Tables, figures, citations, and delivery. latex-empirical-tables builds estout-style LaTeX regression tables with an empirical-economics preamble and repairs common layout failures such as margin overflow, cramped notes, and broken multicolumn rules. stata-style-figures gives matplotlib charts the visual conventions of Stata's stcolor scheme so figures remain consistent across a paper. cite-placement inserts references from a user-supplied, verified bibliography rather than inventing authorities. latex-to-word converts LaTeX manuscripts into native Word documents while preserving tables, equations, figures, footnotes, and cross-references, and markdown-to-pdf produces polished PDFs while checking that referenced images were actually embedded.
- Independent reproducibility auditing. The flagship adversarial-empirical-review skill closes the loop before submission. It re-runs an empirical paper's pipeline in an isolated copy, compares displayed table cells with machine-readable computed values, checks that fresh executions reproduce committed results, and tests whether repeated runs are deterministic. Its strongest mode deletes committed intermediates and rebuilds from raw inputs, so a stale cache or hand-edited table cell cannot survive merely because the code looks plausible. Failed checks enter an adversarial drafter-reviewer-judge process, but the workflow remains fail-closed: it cannot certify a table whose provenance has not been mechanically verified.
- legal-scholarship-skills Claude Code and Codex
A collection of eight Claude Code and Codex skills for the distinctive production pipeline of legal scholarship. Legal articles pose problems that general writing tools rarely handle well: much of the argument is carried in footnotes, citation form is governed by demanding and jurisdiction-specific manuals, authorities must never be fabricated, and even authors who draft in LaTeX eventually exchange redlined Word files with editors and co-authors. The repository covers the path from the first article plan through literature discovery, citation placement, editorial revision, and final document delivery. Each skill installs independently, while the full collection forms an end-to-end law-review toolkit.
- Planning and structured drafting. writing-article-plans interviews the author to settle the thesis, contribution, article type, audience, and structure before prose generation begins. It decomposes each section into paragraph-sized argumentative moves, identifies the function of each paragraph, and marks claims requiring support with typed citation slots instead of invented references. The resulting plan can drive a structured LaTeX drafting process while preserving the author's control over the argument and making source requirements visible before citations are inserted.
- Literature discovery, retrieval, and reading. lit-review-orchestrator starts from a manuscript, abstract, proposal, or research question and returns a deduplicated, relevance-ranked bibliography in which every retained record has been checked against a scholarly index. download-gated-pdfs retrieves the actual PDF binary from publisher, repository, or think-tank pages that serve bot checks and HTML interstitials to non-browser clients. pdf then extracts and analyzes acquired documents, including a footnote-aware route for law-review articles that places each note near its reference point instead of appending a scrambled block of footnotes to the end. Together, the skills preserve an auditable path from discovery to source text.
- Verified citation placement and restyling. The flagship cite-placement workflow uses only references from an author's screened bibliography, verifies them against OpenAlex and Crossref, and inserts them into a new copy of a LaTeX or Word manuscript so the original remains untouched. It supports inline author-date citations and full footnotes under Bluebook, OSCOLA, Chicago, APA, or McGill conventions. Short forms such as Id., supra note, ibid, and OSCOLA cross-references are generated and repairable after manual editing. Case citations, legislation, signals, and discursive footnotes are deliberately preserved rather than mechanically transformed, and a changelog records each decision. A standalone Windows application exposes the restyling pipeline to colleagues who use neither agent.
- Word-centred editorial revision. word-docx reads the redlined Word document returned by an editor, exposes comments and tracked changes for review, prepares response-to-comments documents, and applies approved replies as tracked edits rather than flattening the revision history. latex-to-word supports the round trip for authors who draft in LaTeX: convert an editor's Word file back into workable source, revise it, and return a Word document while checking that footnotes have not silently disappeared. For final delivery, its one-way fidelity engine creates native Word tables, equations, footnotes, figures, and resolved cross-references instead of screenshots or fragile conversions.
- Document delivery and safeguards. markdown-to-pdf turns research memos, cover letters, reports, and other Markdown materials into polished PDFs, with checks that images and linked assets are genuinely present in the output. Across all eight skills, the emphasis is on traceability and safe boundaries: citations come from screened inputs, bibliographic records are verified, originals are preserved, style transformations are logged, footnote counts are checked, and editor-facing changes remain reviewable. Researchers can install the collection as a unified legal-scholarship toolkit or copy only the individual workflow they need, with installation and invocation instructions for both Claude Code and Codex.
- lit-review-orchestrator Python · Claude Code and Codex
A document-driven literature-review pipeline that turns a manuscript, abstract, or proposal into a deduplicated, relevance-ranked bibliography. It extracts a search plan from the document, runs several deep-research engines concurrently (Undermind, Gemini Deep Research, and Google Scholar), and fuses their results into a single list. Before any reference reaches the output, a verification stage confirms each candidate against OpenAlex, Crossref, or Semantic Scholar and drops anything none can confirm, removing the fabricated citations that language-model searches are prone to produce. Distributed as a Claude Code and Codex skill, it delivers the screened results as a relevance-ranked spreadsheet alongside RIS and BibTeX exports.