Skip to content

Latest commit

 

History

History
59 lines (34 loc) · 3.43 KB

File metadata and controls

59 lines (34 loc) · 3.43 KB

Design Principles

The opinions baked into AER-skills.

1. Identification Before Prose

A weak identification strategy cannot be rescued by writing. The skill order enforces this: aer-identification runs before aer-introduction. If the design fails, the contribution sentence cannot survive.

2. One Contribution Per Paper

AER editors desk-reject competent extensions. The aer-topic-selection skill forces the contribution into a single sentence. If the sentence requires "and we also explore ...", that and is where the paper dies.

3. Cross-Subfield Interest Is the Top-5 Filter

The AER bar is not technical quality. It is whether economists outside the author's subfield will cite the paper. Many strong subfield papers belong in the AEJ family or a top field journal. AER-skills makes this routing explicit.

4. Modern Econometrics, Not 1990s Defaults

The skill stack assumes 2026 best practice:

  • Staggered DiD → Callaway-Sant'Anna or Borusyak-Jaravel-Spiess, not TWFE
  • Weak IV → Anderson-Rubin confidence sets, not first-stage F > 10
  • RDD → local linear with MSE-optimal bandwidth, not polynomials of order 4
  • SCM → permutation inference and placebo grid, not just visual fit
  • Shift-share → explicit choice between exogenous-shares and exogenous-shocks identification

5. The Replication Package Is Part of the Paper

The AEA Data Editor's review is enforced and gates publication. aer-replication runs early, not after acceptance. README-first development eliminates retrofitting cost.

6. Editor Time Is the Scarcest Resource

The desk-rejection rate at AER is high. Every formatting, length, and clarity rule in the skill stack is designed to make the first 10 minutes of editor review as efficient as possible:

  • ≤ 100-word abstract
  • No "Introduction" heading
  • ≤ 200-word cover letter
  • Tables that read in one glance
  • Response letters where action and location are visible in the first sentence

7. Anticipate, Don't React

aer-robustness runs before submission to pre-empt the three demands every referee makes (robustness, heterogeneity, mechanism). Reactive revision after the report is twice as expensive and twice as slow.

8. The AEJ Family Is Not a Consolation Prize

Routing a subfield-bounded paper to AEJ:Applied or AEJ:Policy from the start is faster, cheaper, and produces equivalent visibility for many topics. Pride-driven AER submissions cost ~6 months of life per round.

9. Skills Are Routers, Not Replacements

Each skill in this repository solves one slice. aer-workflow exists to compose them. None of them replaces the user's judgment about what makes a good economics paper.

10. English-First, Globally Usable

All skill content is in English so the agent can apply it regardless of the user's primary language. The default landing page README.md is in Chinese for discoverability, with a full English version at README.en.md.

11. Self-Verifying Gates

Every skill ends with a binary, checkable gate — a checklist or a Handoff contract — that the agent must satisfy before declaring the stage done. The skills are self-verifying, not merely advisory. This is the same validation-gate discipline that text-space skill optimizers such as Microsoft SkillOpt use to accept only strictly-improving edits, brought inside each skill so an agent can apply it without an external evaluation harness. See docs/skillopt-evaluation-protocol.md for how these gates feed the held-out evaluation.