name	论文workflow
description	经济学实证论文 9 阶段标准工作流。Use when user starts a new empirical economics paper (course / dissertation / journal). Triggers on /论文workflow, "paper workflow", "empirical econ paper", or descriptions of starting an econ paper task with deadline + word count.

论文workflow — Empirical Economics Paper Workflow

This is the Anthropic Skill version of this workflow, for use with Claude Code's Skill tool, the Claude Agent SDK, or any platform that supports the Anthropic skill format.

For the Codex / generic agent version, see AGENTS.md. For the Claude Code slash command version, see commands/论文workflow.md.

All three files describe the same workflow with the same 9 phases. Pick whichever your platform reads natively.

Activation

When invoked, do not improvise. Follow the 9 phases below in order. Confirm context first (Phase 0).

Phase 0 — Context confirmation (MANDATORY)

Ask the user these 7 questions verbatim. Do not assume answers:

1. 论文类型：课程论文 / 毕业论文 / 期刊投稿？
2. 字数范围 + 截止日期？
3. 主题状态：
   (a) 从零选题（我给你 5 个候选）
   (b) 已定主题（你告诉我）
   (c) 续写已有项目（在哪个目录？）
4. 引用格式：Harvard Manchester / APA 7 / AER 传统 / 其他？
5. 计量方法：你课程教过哪些（DID / IV / RDD / FE / 其他）？
6. 数据约束：必须公开数据？还是可以用机构访问？
7. 工具栈：R / Stata / Python？

If user says "按之前的" / "same as last time", apply defaults:

Harvard Manchester citations
R + public data
~/Desktop/econ-papers/<paper-name>/
TWFE event study + Callaway-Sant'Anna robust

Phase 1 — Topic + econometric setting (~30 min)

Brainstorm 5 candidate topics (each must satisfy: ≤1 day data, ≤3 day identification, anchor paper exists, scope-tractable, R packages mature). Output comparison table. Lock topic. Generate daily-milestone calendar in docs/WORKFLOW.md.

Phase 2 — Literature review (~1.5 h)

Download anchor + critique PDFs. pdftotext -layout + grep appendix tables. Identify 5+ related papers. Write Notes/lit_review.md.

Phase 3 — Data acquisition (~1 h)

Public direct links first. Try R-package hardcoded URLs. Hardcode anchor's appendix parameters. Document every URL in Notes/DATA_ACQUISITION.md. Avoid 24-48h registration loops.

Phase 4 — R analysis (~2.5 h)

6 R scripts: 00_setup.R, 01_clean_<source>.R, 02_compute_distance.R, 03_main_regression.R, 04_robustness.R, 05_tables_figures.R. Use fixest + did. Wrap encoding-prone reads in tryCatch with latin1/windows-1252 fallback. For invalid GADM polygons: st_make_valid() + sf_use_s2(FALSE).

Preferred-spec selection: NEVER use max |t|. Use ex ante criterion (sample median, pre-registered, theoretical motivation).

Default specification: TWFE event study + region FE + country×round FE + Callaway-Sant'Anna robust + 300-permutation placebo + heterogeneity (gender/urban/age) + exogenous-shock interaction.

Phase 5 — Writing (~2 h)

Markdown source + YAML frontmatter → Pandoc → docx. Don't write LaTeX directly. Don't double-number sections. Structure: IMRAD + Lit Review.

Phase 6 — Multi-layer review (~1 h)

Three independent passes:

Content (OVB, identification threats) via claesbackman/AI-research-feedback 2-agent
Format (missing citations, double numbering, broken refs) via oceangis/skill_academic-writing-skills
References (DOI verification per cited paper) → Paper/references_verified.csv

Phase 7 — Word-count adjustment (~30 min)

Per-section count via awk. Each added paragraph must contain ≥1 of: number, citation, mechanism, limitation, future work. No filler.

Phase 8 — Reference-format adaptation (~30 min)

Apply target style. Harvard Manchester defaults:

Single quotes around article titles
pp. prefix for page ranges
doi: lowercase
Available at: for grey literature
(Accessed: <date>) for websites

Phase 9 — Version archive

Immutable Paper/full_paper_v1.md → v9.md. Never overwrite. Generate .docx per version.

Departures from default

Long-form (≥10,000 words): expand §3 descriptives, §4 derivations, §5 sub-sample/mechanism, §6 external validity, full robustness appendix
Short brief (<3,000 words): collapse phases 5-9 into single editorial pass
Non-DID (RDD/IV/RCT replication): Phase 4 changes; ask user about identification setup first

Anti-patterns (never do)

Skip Phase 0 questions
Use max |t| for preferred spec
Fabricate numbers
Pad word count with filler
Write LaTeX directly
Double-number sections
Skip reference DOI verification
Overwrite full_paper_vN.md

论文workflow — Empirical Economics Paper Workflow

This is the Anthropic Skill version of this workflow, for use with Claude Code's Skill tool, the Claude Agent SDK, or any platform that supports the Anthropic skill format.

For the Codex / generic agent version, see AGENTS.md. For the Claude Code slash command version, see commands/论文workflow.md.

All three files describe the same workflow with the same 9 phases. Pick whichever your platform reads natively.

Activation

When invoked, do not improvise. Follow the 9 phases below in order. Confirm context first (Phase 0).

Phase 0 — Context confirmation (MANDATORY)

Ask the user these 7 questions verbatim. Do not assume answers:

1. 论文类型：课程论文 / 毕业论文 / 期刊投稿？
2. 字数范围 + 截止日期？
3. 主题状态：
   (a) 从零选题（我给你 5 个候选）
   (b) 已定主题（你告诉我）
   (c) 续写已有项目（在哪个目录？）
4. 引用格式：Harvard Manchester / APA 7 / AER 传统 / 其他？
5. 计量方法：你课程教过哪些（DID / IV / RDD / FE / 其他）？
6. 数据约束：必须公开数据？还是可以用机构访问？
7. 工具栈：R / Stata / Python？

If user says "按之前的" / "same as last time", apply defaults:

Harvard Manchester citations
R + public data
~/Desktop/econ-papers/<paper-name>/
TWFE event study + Callaway-Sant'Anna robust

Phase 1 — Topic + econometric setting (~30 min)

Phase 2 — Literature review (~1.5 h)

Download anchor + critique PDFs. pdftotext -layout + grep appendix tables. Identify 5+ related papers. Write Notes/lit_review.md.

Phase 3 — Data acquisition (~1 h)

Public direct links first. Try R-package hardcoded URLs. Hardcode anchor's appendix parameters. Document every URL in Notes/DATA_ACQUISITION.md. Avoid 24-48h registration loops.

Phase 4 — R analysis (~2.5 h)

Preferred-spec selection: NEVER use max |t|. Use ex ante criterion (sample median, pre-registered, theoretical motivation).

Default specification: TWFE event study + region FE + country×round FE + Callaway-Sant'Anna robust + 300-permutation placebo + heterogeneity (gender/urban/age) + exogenous-shock interaction.

Phase 5 — Writing (~2 h)

Markdown source + YAML frontmatter → Pandoc → docx. Don't write LaTeX directly. Don't double-number sections. Structure: IMRAD + Lit Review.

Phase 6 — Multi-layer review (~1 h)

Three independent passes:

Content (OVB, identification threats) via claesbackman/AI-research-feedback 2-agent
Format (missing citations, double numbering, broken refs) via oceangis/skill_academic-writing-skills
References (DOI verification per cited paper) → Paper/references_verified.csv

Phase 7 — Word-count adjustment (~30 min)

Per-section count via awk. Each added paragraph must contain ≥1 of: number, citation, mechanism, limitation, future work. No filler.

Phase 8 — Reference-format adaptation (~30 min)

Apply target style. Harvard Manchester defaults:

Single quotes around article titles
pp. prefix for page ranges
doi: lowercase
Available at: for grey literature
(Accessed: <date>) for websites

Phase 9 — Version archive

Immutable Paper/full_paper_v1.md → v9.md. Never overwrite. Generate .docx per version.

Departures from default

Long-form (≥10,000 words): expand §3 descriptives, §4 derivations, §5 sub-sample/mechanism, §6 external validity, full robustness appendix
Short brief (<3,000 words): collapse phases 5-9 into single editorial pass
Non-DID (RDD/IV/RCT replication): Phase 4 changes; ask user about identification setup first

Anti-patterns (never do)

Skip Phase 0 questions
Use max |t| for preferred spec
Fabricate numbers
Pad word count with filler
Write LaTeX directly
Double-number sections
Skip reference DOI verification
Overwrite full_paper_vN.md

workflow

论文workflow — Empirical Economics Paper Workflow

Activation

Phase 0 — Context confirmation (MANDATORY)

Phase 1 — Topic + econometric setting (~30 min)

Phase 2 — Literature review (~1.5 h)

Phase 3 — Data acquisition (~1 h)

Phase 4 — R analysis (~2.5 h)

Phase 5 — Writing (~2 h)

Phase 6 — Multi-layer review (~1 h)

Phase 7 — Word-count adjustment (~30 min)

Phase 8 — Reference-format adaptation (~30 min)

Phase 9 — Version archive

Departures from default

Anti-patterns (never do)

See also

论文workflow — Empirical Economics Paper Workflow

Activation

Phase 0 — Context confirmation (MANDATORY)

Phase 1 — Topic + econometric setting (~30 min)

Phase 2 — Literature review (~1.5 h)

Phase 3 — Data acquisition (~1 h)

Phase 4 — R analysis (~2.5 h)

Phase 5 — Writing (~2 h)

Phase 6 — Multi-layer review (~1 h)

Phase 7 — Word-count adjustment (~30 min)

Phase 8 — Reference-format adaptation (~30 min)

Phase 9 — Version archive

Departures from default

Anti-patterns (never do)

See also