finance-accounting-methods-advisor

Imported from xashgg/finance-accounting-methods-advisor (AGENTS.md). Install upstream with npx skills add xashgg/finance-accounting-methods-advisor. Copyright stays with the author.

Project Guide for Agents

Purpose

This repository packages the finance-accounting-methods-advisor Codex skill. It combines routing instructions, methodological references, reusable research templates, starter Python pipelines, and behavior-focused evaluation cases.

Source-of-truth map

• SKILL.md defines activation, routing, guardrails, and response behavior.
• references/ contains the detailed methodological guidance loaded selectively.
• templates/ contains reusable research artifacts; preserve bracketed unknowns rather than inventing sample sizes, validation results, or model performance.
• scripts/ contains conservative starter pipelines, not turnkey research designs.
• evals/evaluation-cases.md defines expected behavior and regression checks.
• README.md and README.zh-CN.md are synchronized English and Simplified Chinese installation and usage overviews.

Maintenance rules

• Keep the workflow construct-first: construct, unit, task, baseline, preferred method, validation, implementation, and econometric use.
• Preserve the distinction between discovery and production measurement, and between prediction and causal identification.
• Treat LLM output as a measurement instrument that requires human validation.
• Require time-, firm-, document-family-, or group-aware splits when the research setting makes random row splits vulnerable to leakage.
• Keep SKILL.md compact. Put detailed explanations in the relevant reference and route to them from the skill.
• When adding or renaming a reference, template, script, or eval, update the file maps and routing tables in SKILL.md, both README files, and the relevant subdirectory guide in the same change.
• Do not hard-code current package behavior without checking official documentation.
• Do not add dependencies unless the methodological benefit justifies them; update scripts/requirements-starter.txt and scripts/scripts-README.md together.

Verification

Run command-interface smoke checks after script changes:

python -B scripts/init_research_project.py --help
python -B scripts/text_classifier.py --help
python -B scripts/embedding_similarity.py --help
python -B scripts/topic_discovery.py --help
python -B scripts/llm_classification.py --help
python -B scripts/doubleml_template.py --help

After changing SKILL.md or a reference, rerun at least Evals 1, 3, 5, 7, 8, and 10 from evals/evaluation-cases.md. Verify local paths named in Markdown and scan for stale relative-time language before handing off.