kennethkhoocy/applied-micro-skills

Automates verification that empirical research tables are correct by comparing them against the underlying data and analysis code through an adversarial, regression-gated review pipeline. Use when you need to audit whether numbers in LaTeX tables match computed results.

Audit the annotation protocol to ensure the model receives the same input as human annotators, preventing recall ceilings and misdiagnosed failures. Use before designing any label-replication model, or when a validation report shows low recall on label subsets.

Fixes a silent throughput ceiling in asyncio LLM batch pipelines by overriding the default httpx connection pool limit to match the desired concurrency. Use when raising the asyncio semaphore above ~100 produces no gain.

Unified router for placing pre-screened citations into manuscripts or restyling existing citations. Supports inline, footnote placement, and full style conversion for LaTeX and Word documents.

Runs a deep literature search using Google Gemini's Deep Research agent via API, parses the cited report into structured data for a literature review pipeline. Use only when explicitly requested as an API-driven alternative to browser-based deep searches.

Downloads the actual PDF from bot-gated websites (e.g., taxpolicycenter.org, SSRN mirrors) that return HTML instead of PDFs. Uses the Wayback Machine's raw-content (id_) URL to obtain the original binary, bypassing bot challenges.

Computes publication-quality cumulative abnormal returns (CARs) and CAAR with robust event-study test statistics (Patell, BMP, GRANK-T, etc.). Use for event studies, thin-trading adjustment, and market-model regressions.

Performs keyless keyword searches across OpenAlex, Crossref, and Semantic Scholar for literature reviews. Use as a fallback discovery channel when API keys for other search services are unavailable.

Prepares human-readable adjudication sheets from LLM pipeline outputs, recovering full evidence and handling Excel cell limits, multi-line text, and truncation issues to ensure valid human rulings.

Verifies LLM model consistency before resuming a paused batch-scoring campaign by comparing fresh responses to cached outputs. Detects silent model drift from alias repoints or serving revisions to prevent splicing different model versions into one measure.

Check if an LLM classifier's validation failure is gold-bound before attempting prompt fixes, and use a gated pilot design to verify improvements when not gold-bound.

Set up, format, fix, and clean up LaTeX regression tables for empirical economics and finance papers. Handles estout fragments, margin overflow, alignment, significance stars, and Misplaced \omit errors.

Convert between LaTeX and Microsoft Word for academic manuscripts, preserving tables, equations, and cross-references, and assemble .tex from PDFs, .docx, and LLM-generated content.

Orchestrates a multi-source literature search pipeline from a document. Give it a .tex or .docx describing an article, extracts a search plan, runs Undermind and Google Scholar searches, then merges, deduplicates, and screens results.

Screens paper abstracts against a research prompt, producing a relevance score (1-10), paper type, methodology, and relationship to the user's work. Part of a literature review pipeline, used when explicitly requested.

Merge and deduplicate papers from literature review pipeline stages using exact DOI matching and LLM fuzzy matching. Produces a master list with provenance tracking for research workflows.

Convert Markdown files to polished PDFs while preserving all images, tables, code blocks, and table-of-contents links. Ideal for turning GitHub-flavored Markdown documents (READMEs, reports, notes) into PDF deliverables.

Diagnose and fix slow pyfixest regression grids by sharding the model list across OS processes. Use this skill when a loop of pyfixest models on large panels is CPU-bound despite GPU assignment.

When adding 'cupy64' backend to pyfixest, fully absorbed regressors are not dropped, producing garbage coefficient rows. This skill guides verification: diff outputs, ignore absurd estimates, and ensure headline coefficients remain correct.

Automates a Google Scholar Labs deep search from a research question, scraping and enriching paper citations via BibTeX export. Used in Stage 2 of a literature review pipeline when explicitly requested.

Enables execution of Stata commands and do-files through Pystata for econometric analysis, data exchange with pandas, and retrieval of stored results without intermediate files. Use for regressions, IV, diff-in-diff, RDD, and more.

Applies the Stata 18/19 default (stcolor) scheme to matplotlib figures, embedding Arial fonts, white background, recessive grid, and validated blue/red/gray palette. Use when generating publication-quality charts for papers, reports, or slides.

Searches academic databases (Google Scholar, SSRN, NBER, Semantic Scholar, etc.) for working papers, citation chains, and forthcoming articles, using LLM-based query condensation. Designed as Stage 4 of a lit review pipeline.

Automates deep literature search on Undermind.ai via headless browser, then parses and enriches references with DOIs and abstracts for a literature review pipeline. Use when integrating Undermind searches into automated academic workflows.

A keyless, agent-driven web search channel for discovering academic papers using Claude Code's WebSearch and WebFetch tools, intended as a fallback when no search API credentials are available.