Skip to content

EconAgentSkills

external · status: dormant · focus: analysis · discipline: economics · started: 2026

Project page: https://github.com/JonasWeinert/EconAgentSkills

Licence: CC0-1.0

Source: projects/landscape/econ-agent-skills.yml

Positioning

Ten opinionated agent skills covering the applied-econometrics workbench in four stages: data (working-with-data, api-data-fetcher, stata-data-cleaning), analysis (r-econometrics, stata-regression, python-panel-data), writing (academic-paper-writer, latex-tables) and communication (econ-visualization, beamer-presentation). Each ships a SKILL.md routing layer plus reference.md and runnable examples in R, Python, Stata, Julia and LaTeX. Sits in the RISE analysis layer beside statspai, stata-code and stata-mcp, but supplies method guidance to a coding agent rather than an execution surface.

Distinctive contribution

The only pack in the catalog that pins the modern heterogeneity-robust estimators by name rather than gesturing at "econometrics": Callaway-Sant'Anna, Sun-Abraham, Borusyak-Jaravel-Spiess and de Chaisemartin-D'Haultfoeuille for staggered DiD, weak-IV-robust inference with first-stage F-tests and Anderson-Rubin confidence sets, rdrobust optimal bandwidths, and wild cluster bootstrap — written explicitly to stop an agent reaching for plain TWFE on staggered treatment. Its second contribution is AGENT_POLICY.md, a four-level escalation ladder (ASK = block and present options; DEFAULT+flag = use the stated default and notify; DOCUMENT = record the assumption in a decisions log and continue; PROCEED = execute) with every generated script carrying a decisions-log header whose uncertain items are tagged "[HUMAN: please confirm]".

Evaluation scores

Dimension Score (0–3) Note
Lifecycle coverage 2 Four stages — API data fetching, panel/DiD/IV/RDD analysis, script generation, and LaTeX table + manuscript writing — with large gaps: no literature discovery or synthesis, no referee simulation, no replication.
Autonomy level 1 Copilot by explicit policy: the ASK level of AGENT_POLICY.md blocks and presents options without proceeding, and every generated script carries a decisions-log header with '[HUMAN: please confirm]' items — human approval is designed into each significant methodological choice.
Architectural transparency 3 Every SKILL.md, its reference.md, the runnable R/Python/Stata/Julia/LaTeX examples and AGENT_POLICY.md are in the repo; the README states the CC0 dedication and documents the routing and decision-tree design. No evaluation harness exists to publish.
Inputs supported 2 Multiple input forms (user datasets in several formats, a free-text analysis request, an existing draft for latex-tables / academic-paper-writer) plus data-source access via api-data-fetcher; no literature corpus or bibliographic connector of any kind.
Outputs / reproducibility 2 Produces runnable analysis scripts, LaTeX tables and vector figures, and mandates a decisions-log header plus DIME master-do-file discipline — genuine provenance for the choices made — but there is no end-to-end reproducibility guarantee and no manifest tying outputs back to inputs.
Internal evaluation 0 No evaluation, test suite, or worked-benchmark comparison reported; the method choices are asserted from the literature, not validated against a gold standard. Verified by reading the README on 2026-09-08.
Openness 2 A LICENSE file in the repo confirms CC0 1.0 Universal (maximally permissive; re-verified 2026-10-01, resolving the earlier GitHub-API NOASSERTION ambiguity), and the skills are plain markdown with runnable examples, but the three Stata skills additionally require a commercial Stata license to reproduce, so not all examples are commodity-hardware reproducible.
Maturity / traction 1 24 stars, 3 forks, still only 3 commits — all on 2026-05-05, the day the repo was created, with nothing pushed since (confirmed 2026-10-01). A well-made single-drop personal pack, not a maintained project.
Cross-family policy 0 No cross-family review mechanism: the skills are runtime-portable markdown but say nothing about which model executes versus which reviews, and no second-family check is defined anywhere. Portability is not a cross-family policy.
Runtime assurance 1 One light in-flight mechanism, applied consistently: the ASK/DEFAULT+flag/DOCUMENT/PROCEED ladder plus the decisions-log header surface and escalate methodological assumptions. Nothing verifies the numbers — no result audit, no re-estimation check, no figure or claim inspection.
Cross-platform portability 2 Built on the open SKILL.md standard with a Python installer targeting Claude Code, Cursor and Codex, and examples in five languages — 3 documented runtimes, short of the 5+ environments band 3 requires.

Scored on 2026-10-01. See the evaluation rubric.

Tags

Pipeline stages: data-acquisition data-analysis code-generation paper-drafting

Architectural features: human-in-loop tool-use

Inputs: user-dataset analysis-request paper-draft

Outputs: analysis-code latex-tables figures decisions-log beamer-slides

Data sources: user-provided public-apis

Knowledge sources: modern-did-iv-rdd-literature dime-analytics-conventions

Limitations

  • Dormant: 3 commits, all on the creation date 2026-05-05, none in the five months since (confirmed 2026-10-01).
  • License confirmed as CC0 1.0 Universal via the repo's LICENSE file (earlier GitHub-API NOASSERTION reading was a parsing artifact, not a real ambiguity).
  • Deliberately opinionated: the defaults are one defensible position in live methodological debates, and a research design the author did not anticipate will fight the skill rather than be served by it.
  • Method coverage is a mid-2026 snapshot; a named-estimator pack ages faster than a generic one, and there is no maintainer activity to refresh it.
  • No evaluation of whether following the skills produces better estimates than an unguided agent.
  • The three Stata skills are unusable without a commercial Stata license.
  • No literature, review, or replication coverage — it must be paired with a review pack to cover a paper's lifecycle.