📌 A Codex skill for systematic, auditable, and interpretation-first specification search in empirical economics.
xianzhu-skill is built for a very common research situation:
the question is already fixed, but the regression result is still not stable enough, not convincing enough, or not yet significant enough to write into the paper.
Instead of encouraging blind p-hacking, this skill turns “keep trying around the same x -> y relationship” into a disciplined empirical workflow with:
- clear boundaries 🧭
- admissible transformations 🔁
- transparent experiment logs 🗂️
- model-selection discipline 📐
- manuscript-facing output logic 📝
Use xianzhu-skill when the user says things like:
- “试到显著”
- “换个口径试试”
- “这个变量能不能做
log/asinh/ln(x+1)？” - “方向是对的，但不显著，怎么办？”
- “别换研究问题，围绕当前方程继续试”
Typical use cases:
- improving the stability of an existing empirical result
- exploring admissible transformations of the current treatment or outcome
- deciding whether to keep the current proxy or move to a substitute measure
- preserving a full audit trail of failed and retained specifications
- preparing manuscript-ready justification for why one specification is preferred
This skill is intentionally narrow and method-conscious.
The research question should already be known. This skill is not for inventing topics from scratch.
It first tests transformations of the current variable, and only after that considers same-concept substitutes.
A more significant result is not automatically a better result if its economic meaning becomes weaker.
Failed specifications are not noise to be hidden; they are part of the empirical record.
The skill is designed for real economics projects where do-files, logs, and output tables need to be reproducible and paper-ready.
flowchart TD
A[Lock research question] --> B[Decide whether to search x, y, or both]
B --> C[Choose compatible estimation methods]
C --> D[Search current indicator first]
D --> E[Record every attempt]
E --> F{Current indicator fails?}
F -- No --> G[Recommend retained specification]
F -- Yes --> H[Consider same-concept substitute indicators]
H --> I[Compare interpretability, direction, and significance]
I --> G
In practice, the skill works in two modes:
-
Mode A: identify then search
For cases where the exact variable or estimating equation is not yet fully fixed. -
Mode B: direct experiment
For cases where the user already has a regression equation, result table, or do-file and wants to keep pushing the same empirical question.
xianzhu-skill/
├── SKILL.md
├── README.md
├── LICENSE
├── .gitignore
├── agents/
│ └── openai.yaml
├── evals/
│ └── evals.json
└── references/
├── case-examples.md
├── method-menu.md
├── mode-selection.md
├── spec-log-template.md
├── stata-output-conventions.md
└── transform-playbook.md
The repository includes several supporting reference files:
-
references/method-menu.md
A compact menu of admissible estimation approaches. -
references/transform-playbook.md
Practical guidance onlog,ln(x+1),asinh, winsorization, discretization, and re-expression. -
references/spec-log-template.md
Minimum traceability requirements for specification-search logs. -
references/stata-output-conventions.md
Naming, do-file separation, and output/log conventions for Stata-based empirical work. -
references/case-examples.md
Realistic cases showing how the workflow should behave in practice.
This skill does not:
- invent a research question from nothing
- replace literature grounding
- fabricate substitute variables without support
- hide unsuccessful results
- equate “more significant” with “more publishable”
- treat aggressive p-hacking as valid empirical work
A good run of xianzhu-skill should usually return:
- mode identification (
Mode AorMode B) - the locked-in
x,y, and method choice - the side currently being searched
- a list of tried specifications
- the preferred specification
- rejected specifications and reasons
- a complete experiment log for reproducibility
This repository is especially useful for:
- empirical economics researchers
- paper-revision workflows
- Stata-first regression projects
- Codex users who need a specification-search skill with explicit methodological boundaries
If you want the repository About box to look complete, a good setup would be:
Description
A Codex skill for systematic and auditable specification search in empirical economics.
Website
Leave blank for now, unless you want to point it to a documentation page or personal site.
Topics
codex-skill, economics, empirical-economics, stata, regression, specification-search, research-workflow
MIT