From empirical papers to reproducible agent skills. Agent skills for empirical economics, causal inference, and reproducible research writing.
EmpiriForge turns empirical research practice into reusable AI-agent workflows: research questions, contribution framing, identification strategies, data workflows, regression tables, robustness checks, reproducibility packages, and reviewer-style audits.
中文理解：EmpiriForge 是“实证研究熔炉”。它不是一组泛泛的论文写作提示词，而是把实证经济学研究中的隐性经验锻造成可安装、可审计、可复用的 SKILL.md 工作流。
EmpiriForge focuses on original empirical research, especially applied economics and quantitative social science.
Use it for:
- academic paper production from topic/proposal to manuscript and delivery;
- empirical economics paper planning and revision;
- DID, IV, RDD, event-study, panel FE, synthetic-control, matching, and causal ML design review;
- CATE/ITE/uplift estimation planning, causal-GNN audits, graph-based perturbation prediction, and causal-ML validation;
- machine-learning prediction-model papers with tidymodels, random forest, XGBoost, logistic baselines, PLS/VIP models, stable carbon-emission forecasting, distribution-shift validation, ROC/AUC, calibration, decision curves, and leakage checks;
- bootstrap uncertainty for standard errors, confidence intervals, prediction bands, AUC intervals, and model-robustness checks;
- publication-ready R figures, correlation matrices, corrplot correlograms, journal visualization standards, and reviewer-proof figure audits;
- separating identification credibility from code reproducibility;
- building replication packages and pre-submission audits;
- turning repeated research routines into reusable agent skills.
For systematic reviews, meta-analysis, umbrella reviews, second-order meta-analysis, or AI-assisted evidence synthesis, use the companion project EvidenceForge.
Modern empirical research has two bottlenecks:
- The intellectual bottleneck: clarifying the research question, contribution, estimand, identification assumptions, and interpretation.
- The operational bottleneck: cleaning data, running code, reproducing tables, checking diagnostics, and preparing replication packages.
Large language models are most useful when they sit inside a controlled workflow: they coordinate, inspect, route tasks, draft reports, and keep a research ledger. Deterministic code should still own numerical computation. Human researchers should still own scientific judgment.
This repository is inspired by Scaling Reproducibility: An AI-Assisted Workflow for Large-Scale Replication and Reanalysis: separate orchestration, skill memory, deterministic execution, and human review.
EmpiriForge/
skills/
academic-paper-pipeline/
SKILL.md
references/
agricultural-systems-scope-fit.md
paper-skill-stack.md
templates/
agricultural-systems-scope-fit-audit.md
paper-production-ledger.md
submission-delivery-checklist.md
empirical-research-forge/
SKILL.md
references/
economics-workflow.md
identification-checklist.md
method-source-map.md
reproducibility-package.md
research-transparency.md
templates/
research-audit-report.md
econ-identification-skeptic/
SKILL.md
references/
design-diagnostics.md
identification-decision-rules.md
causal-ml-estimator-selector/
SKILL.md
references/
causal-gnn.md
causal-ml-toolkit.md
templates/
causal-gnn-audit.md
causal-ml-design-card.md
causal-ml-validation-plan.csv
graph-causal-validation-schema.csv
prediction-modeling-forge/
SKILL.md
references/
bootstrap-uncertainty.md
pls-vip-modeling.md
stable-carbon-emission-forecasting.md
tidymodels-prediction-workflow.md
templates/
bootstrap-uncertainty-audit.md
pls-vip-audit.md
prediction-model-audit.md
prediction-model-reporting-checklist.csv
stable-environment-validation-schema.csv
stable-time-series-prediction-audit.md
scripts/
bootstrap_metric_ci.py
pls_vip_template.R
tidymodels_binary_classification_template.R
publication-figure-forge/
SKILL.md
references/
correlation-plot-playbook.md
figure-type-playbook.md
journal-figure-standards.md
templates/
correlation-plot-audit.md
costar-figure-prompt.md
figure-review-checklist.md
scripts/
corrplot_correlation_template.R
publication_theme.R
stata-econometrics-runner/
SKILL.md
references/
pdf-docs-and-manuals.md
stata-replication.md
windows-stata.md
templates/
empirical-project-context.md
stata-run-report.md
scripts/
find-stata.ps1
docs/
method-sources.md
source-crosswalk.md
reading-list.md
Scaling_Reproducibility_中文翻译.md
Scaling_Reproducibility_中文解读.md
Scaling_Reproducibility_可提炼skills.md
examples/
example-prompts.md
demo-audit-report.md
Pipeline skill:
academic-paper-pipeline
Core empirical skill:
empirical-research-forge
Specialized companion skill:
econ-identification-skeptic
Specialized causal-ML skill:
causal-ml-estimator-selector
It also audits causal GNN papers: graph-based therapeutic perturbation prediction, LLM-enhanced GNN mechanism analysis, causal disentangled graphs, and fault-diagnosis causal subgraphs.
Prediction-modeling skill:
prediction-modeling-forge
PLS/VIP, bootstrap uncertainty, and stable prediction are part of prediction-modeling-forge, not separate repositories. Use it when a model paper needs small-sample correlated-predictor modeling, VIP ranking, confidence intervals, prediction bands, uncertainty around AUC or fitted values, or cross-region/cross-industry/cross-policy validation for time-series forecasts such as enterprise carbon emissions.
Publication-figure skill:
publication-figure-forge
It includes a corrplot-based correlation-matrix workflow with nine reusable variants: pie, mixed circle/color, significance labels, ellipse, coefficient labels, lower triangle, circle significance, grayscale shade, and clustered rectangles.
Stata workflow skill:
stata-econometrics-runner
EmpiriForge is method-aware, not method-omniscient. It is grounded in a small set of research traditions:
- AI-assisted reproducibility workflows;
- applied econometrics and causal inference;
- research transparency and replication-package standards;
- practical impact-evaluation resources.
See:
- Method Sources
- Source Crosswalk
- Reading List
- Stata Agent Workflow
- Knowledge Graph Navigation
- Irrigation Expansion Causal-Forest Paradigm
- Empirical Agent Ecosystem Bridge
The repository does not replace econometrics textbooks, journal policies, or human peer review. Its purpose is to make those standards easier to apply repeatedly.
Graphify can be used as an optional navigation layer for large empirical projects, replication packages, and notes folders. It helps an agent start from GRAPH_REPORT.md and focused graph queries before reading raw files. It does not replace EmpiriForge guardrails for identification, reproducibility, prediction validity, Stata execution, or figure review.
See Knowledge Graph Navigation.
Input:
A DID paper draft + regression tables + replication folder
Skills:
empirical-research-forge
econ-identification-skeptic
Output:
Research ledger
Identification audit
Robustness checklist
Reproducibility package review
Reviewer-style action plan
Example prompt:
Use empirical-research-forge and econ-identification-skeptic to audit my DID paper.
Separate code reproducibility from identification credibility.
EmpiriForge 可以理解为：
面向经济学实证论文的 AI 研究工作流：把选题、文献定位、识别策略、数据处理、回归表、稳健性检验、复现包和审稿式批评组织成可复用的 agent skills。
它优先服务中文研究者，但保留英文输出能力，适合双语论文笔记、英文投稿、中文解释和 GitHub 发布。
For Codex:
Copy-Item -Recurse ".\skills\empirical-research-forge" "$env:USERPROFILE\.codex\skills\empirical-research-forge"
Copy-Item -Recurse ".\skills\academic-paper-pipeline" "$env:USERPROFILE\.codex\skills\academic-paper-pipeline"
Copy-Item -Recurse ".\skills\econ-identification-skeptic" "$env:USERPROFILE\.codex\skills\econ-identification-skeptic"
Copy-Item -Recurse ".\skills\causal-ml-estimator-selector" "$env:USERPROFILE\.codex\skills\causal-ml-estimator-selector"
Copy-Item -Recurse ".\skills\prediction-modeling-forge" "$env:USERPROFILE\.codex\skills\prediction-modeling-forge"
Copy-Item -Recurse ".\skills\publication-figure-forge" "$env:USERPROFILE\.codex\skills\publication-figure-forge"
Copy-Item -Recurse ".\skills\stata-econometrics-runner" "$env:USERPROFILE\.codex\skills\stata-econometrics-runner"For Claude Code:
cp -r skills/empirical-research-forge ~/.claude/skills/empirical-research-forge
cp -r skills/academic-paper-pipeline ~/.claude/skills/academic-paper-pipeline
cp -r skills/econ-identification-skeptic ~/.claude/skills/econ-identification-skeptic
cp -r skills/causal-ml-estimator-selector ~/.claude/skills/causal-ml-estimator-selector
cp -r skills/prediction-modeling-forge ~/.claude/skills/prediction-modeling-forge
cp -r skills/publication-figure-forge ~/.claude/skills/publication-figure-forge
cp -r skills/stata-econometrics-runner ~/.claude/skills/stata-econometrics-runnerThen ask your agent:
Use academic-paper-pipeline to move this project from proposal to manuscript and delivery.
For empirical research audits:
Use empirical-research-forge to audit my DID paper draft and replication package.
For Stata projects:
Use stata-econometrics-runner to map this Stata replication package, run the master do-file if Stata is available, and summarize log/output mismatches.
For causal ML projects:
Use causal-ml-estimator-selector to choose between S/T/X/R/DR learners, uplift forests, and IV-style estimators for my treatment-effect project. Build a design card and validation plan.
For graph-based causal ML:
Use causal-ml-estimator-selector to audit this causal-GNN paper. Separate graph semantics, intervention claims, proxy-graph assumptions, validation splits, and safe causal wording.
For publication figures:
Use publication-figure-forge to turn this R data frame into a journal-ready scatterplot/boxplot/heatmap/KM curve. Include reviewer-proof checks and export as TIFF 300 DPI.
For correlation matrices:
Use publication-figure-forge to make a corrplot correlation matrix from input.txt. Check whether rows are features or samples, use Spearman correlation with BH-adjusted p-values, and export a publication-ready PDF.
replication-package-janitor: prepare and repair reproducibility packages.academic-paper-pipeline: topic-manuscript-delivery orchestration. Added.causal-diagnostics-runner: deterministic diagnostics for IV, DID, RDD, and event studies.causal-ml-estimator-selector: CATE/ITE/uplift estimator choice, causal-GNN audit, validation workflow, and boundary checks between treatment-effect estimation and causal-invariant prediction. Added.stata-econometrics-runner: Stata.do/.logworkflow for empirical replication. Added.paper-journalist: structured writing, contribution framing, and revision response.publication-figure-forge: R/ggplot2 and corrplot journal-ready visualization workflow, including correlation matrices and reviewer-proof figure audits. Added.prediction-modeling-forge: prediction modeling plus bootstrap uncertainty, stable time-series prediction, and distribution-shift validation for model-performance intervals and fitted-value uncertainty. Added.- optional deterministic scripts for table matching, code inventory, and output verification.
Companion project:
- EvidenceForge: systematic review, meta-analysis, umbrella review, second-order meta-analysis, and AI-assisted evidence synthesis.
This is an early research-skill scaffold. The first version focuses on process quality, identification discipline, reproducible research, and bilingual research writing. Deterministic analysis scripts can be added later.