Tax Assistant

Workflow

1. Input: The user provides a folder of PDF or CSV documents containing financial information. The folder and file names should provide context about the financial institution and date range.
2. Iteration: The agent iterates through each file alphabetically.
3. Country Context: Before processing any files, review references/country_tax_details.md to understand the applicable tax authority, fiscal year boundaries, local currency, and any country-specific adjustments that must be applied across institutions. If this document is a stub, the agent should ask the user to provide a relevant source for regenerating the document. When this occurs, all spreadsheet stubs in the spreadsheets directory should be updated with country-specific columns, while retaining the generic columns.
4. Institution-Specific Knowledge: For each file, the agent consults a knowledge base of .md files stored in the skill's references folder. Each file corresponds to a specific financial institution and contains instructions for processing its documents.
5. New Institution Reference (Conditional): Before processing a file, try to infer the financial institution name from the folder or filename (e.g. discovery, easy_equities, standard_bank) and check for a corresponding .md reference file in the references folder. If no suitable reference document exists or the existing reference file is a stub for that institution, the agent must create/elaborate the markdown reference file in references (using a clear, institution-specific structure with acronym_institution_name.md). Use template.md as a base and outline how to normalize that institution's statements for the fiscal year and tax fields defined in references/country_tax_details.md.
6. PDF Processing: If the file is a PDF, the agent uses the pdf-vision-ocr skill to convert it to a .md file, providing a generic OCR prompt that asks for faithful extraction of all text and tables.
7. Data Extraction: The agent inspects the generated .md file and applies the institution-specific knowledge from the relevant reference document to locate and interpret the target data fields.
8. Normalization: When necessary, normalize the extracted data to the official currency and fiscal year dates described in references/country_tax_details.md. This may involve consulting references/forex_rates.csv (or equivalent) when conversions are required.
9. Spreadsheet Update: Update the relevant .csv stub spreadsheet with the actual figures. If a spreadsheet is lacking for a relevant tax code or field, create a new stub spreadsheet.
10. Statement of Assets & Liabilities Update: Whenever possible, update the market value of any assets as at the end of the fiscal year in spreadsheets\statement_of_assets_and_liabilities.csv. IMPORTANT: Review the local tax authority's guidance (see references/country_tax_details.md for the authoritative source) before updating this file.
11. Auditing: Add file and line number references for each entry for auditing purposes. All file references must point to the generated .md files, not the original .pdf files.
12. File Completion: Move on to the next file until all files have been processed.
13. Final Audit: After all files have been processed, conduct a comprehensive audit of all spreadsheets to ensure the data is accurate and that all entries have the correct file and line number references.
14. Process Review: After completing the final audit, write up a review.md document noting any issues, ambiguities or insights that may have arisen while processing the documents and completing the tax spreadsheets. Conclude with recommendations for improving the SKILL.md file or any other skill guidance documentation.