IPO申报材料复核

Use this skill when the user asks to review IPO filing materials, inquiry responses, prospectuses, audit reports, financial-statement notes, or related disclosure files.

Bundled Tool Location

This skill is self-contained when installed under Codex user skills. The bundled tool directory is the skill folder itself:

%USERPROFILE%\.codex\skills\ipo-review

If the current working directory contains main.py, src/, and config/, use the current directory as the tool project. Otherwise, use this skill folder as the tool project. Do not rely on the original desktop development folder being present.

Preflight

1. Confirm the selected tool project contains:
   • main.py
   • src/
   • config/
   • input/
   • output/
2. Confirm input/ contains materials to review. Supported formats are PDF, DOCX, XLSX, and TXT.
3. If dependencies are missing during execution, install from the tool project with:

pip install -r requirements.txt

The review program itself must run locally. Do not upload source materials or call external APIs.

Run

From the selected tool project directory, execute:

python main.py

The program reads input/, writes output/, and generates:

output/IPO问询回复复核分析报告.html
output/issues.json
output/financial_facts.csv
output/evidence_index.json
output/comparison_exclusions.csv
output/arithmetic_skips.csv
output/parse_errors.csv
output/run_log.txt

Report Interpretation

After running, read output/issues.json and output/run_log.txt.

Prioritize issues where:

review_priority == "key"

These are the "待人工复核（重点）" items. For each key item, summarize:

• issue id, category, level, and priority;
• source files and evidence positions;
• source values and original text snippets when available;
• conclusion, caliber analysis, and suggested manual check;
• whether the item is likely a real issue or a candidate requiring manual confirmation.

For review_priority == "normal" and review_priority == "noise", provide counts and category summaries by default. Expand details only if the user asks.

Diagnostics

Use these files for quality checks:

• comparison_exclusions.csv: facts not compared and why.
• arithmetic_skips.csv: tables skipped for arithmetic checks and why.
• parse_errors.csv: real parsing failures.
• financial_facts.csv: extracted facts, units, periods, scopes, and metrics.
• evidence_index.json: paragraph/table-cell evidence locations.

When assessing whether the tool is reliable, check:

• whether parse errors are zero or explainable;
• whether key issues are few enough for manual review;
• whether high-priority items have usable evidence locations;
• whether obvious noise is labeled noise instead of key.

Required Caveats

Always tell the user:

• This is a review assistant, not a substitute for professional judgment.
• It is strongest at numeric checks: cross-file values, units, periods, and table arithmetic.
• It does not understand all narrative or semantic contradictions.
• "No issue reported" does not mean the filing is confirmed correct.
• The user should still perform manual reading and professional review.

Do not overstate certainty. Do not describe unreported areas as verified.