Tax Filing Skill

Prepare federal and state income tax returns: read source documents, compute taxes, fill official PDF forms.

Year-agnostic — always look up current-year brackets, deductions, and credits. Never reuse prior-year values.

Folder Structure

Organize all work into subfolders of the working directory:

working_dir/
  source/              ← user's source documents (W-2, 1099s, prior return, CSVs)
  work/                ← ALL intermediate files (extracted data, field maps, computations)
    tax_data.txt       ← extracted figures from source docs
    computations.txt   ← all tax math (federal, state, capital gains)
    f1040_fields.json  ← field discovery dumps
    f8949_fields.json
    f1040sd_fields.json
    ca540_fields.json
    expected_*.json    ← verification expected values
  forms/               ← blank downloaded PDF forms
    f1040_blank.pdf
    f8949_blank.pdf
    f1040sd_blank.pdf
    ca540_blank.pdf
  output/              ← final filled PDFs + fill script
    fill_YEAR.py       ← the fill script
    f1040_filled.pdf
    f8949_filled.pdf
    f1040sd_filled.pdf
    ca540_filled.pdf

Create these folders at the start. Keep the working directory clean — no loose files.

Context Budget Rules

These rules prevent context blowouts that cause compaction:

NEVER read PDFs with the Read tool. Each page becomes ~250KB of base64 images (a 9-page return = 1.8 MB). Extract text instead:

python3 -c "
import pdfplumber
with pdfplumber.open('source/document.pdf') as pdf:
    for p in pdf.pages: print(p.extract_text())
"

NEVER read the same document twice. Save extracted figures to work/tax_data.txt on first read.
Run field discovery ONCE per form as a bulk JSON dump to work/. Do NOT use --search repeatedly.
Save all computed values to work/computations.txt so they survive compaction.

Workflow

Step 1: Gather Source Documents

Ask the user what documents they have. Read files from source/ (move them there if needed). Use pdfplumber for PDFs, Read tool for CSVs.

Save all extracted figures to work/tax_data.txt immediately — one section per document with every relevant number.

Step 2: Confirm Filing Details — MANDATORY

You MUST ask the user every one of these questions and WAIT for answers before proceeding. Do NOT skip this step even if you think you know the answers from memory or source documents. Tax returns are legal documents.

Filing status (Single, MFJ, MFS, HOH, QSS)
Dependents (number, names)
State of residence
Standard vs. itemized deduction preference
Digital asset / cryptocurrency transactions (Yes/No) — stock trades are NOT digital assets
Health coverage status (for CA)
Any estimated tax payments made
Any other credits or adjustments

Do NOT proceed to Step 3 until the user has answered. "Same as last year" counts as confirmation.

Step 3: Look Up Year-Specific Values

Research from IRS.gov and FTB.ca.gov:

Federal tax brackets, standard deduction, QDCG 0%/15%/20% thresholds
State tax brackets, standard deduction, personal exemption credit

Save to work/computations.txt.

Step 4: Compute Federal Return

Gross Income: W-2 wages (1a) + interest (2b) + dividends (3b) + capital gain/loss (7)
Adjustments → AGI (Line 11)
Deductions → Taxable Income (Line 15)
Tax: use QDCG worksheet if qualified dividends/capital gains exist
Credits, other taxes → Total Tax (Line 24)
Payments (withholding, estimated) → Refund/Owed
If refund: collect direct deposit info (routing, account, type)

Save all line values to work/computations.txt.

Step 5: Compute Capital Gains (if applicable)

Form 8949: individual transactions (Part I short-term, Part II long-term)
Schedule D: totals, $3,000 loss limitation, carryover calculation
Net gain/loss → 1040 Line 7

Step 6: Compute State Return (CA Form 540)

Federal AGI → CA adjustments → CA taxable income
Tax from brackets − exemption credits → total tax
Withholding → Refund/Owed

Step 7: Download Blank PDF Forms

Save to forms/ directory.

IRS: Use /irs-prior/ for prior-year forms (/irs-pdf/ is always current year):

https://www.irs.gov/pub/irs-prior/f1040--YEAR.pdf
https://www.irs.gov/pub/irs-prior/f8949--YEAR.pdf
https://www.irs.gov/pub/irs-prior/f1040sd--YEAR.pdf

CA: ftb.ca.gov/forms/YEAR/ for state forms.

Verify each download has %PDF- header (not an HTML error page).

Step 8: Discover Field Names & Fill Forms

Discovery — ONCE per form, use `--compact`

python scripts/discover_fields.py forms/f1040_blank.pdf --compact > work/f1040_fields.json
python scripts/discover_fields.py forms/f8949_blank.pdf --compact > work/f8949_fields.json
python scripts/discover_fields.py forms/f1040sd_blank.pdf --compact > work/f1040sd_fields.json
python scripts/discover_fields.py forms/ca540_blank.pdf --compact > work/ca540_fields.json

--compact outputs a minimal {field_name: description} mapping — each field name is paired with its tooltip/speak description so you can map line numbers to field names directly without manual inspection. Radio buttons include their option values (e.g. {"/2": "Single", "/1": "MFJ"}).

Do NOT use --search repeatedly or --json (which dumps raw metadata and wastes context).

HARD FAIL: If discovery returns 0 human-readable descriptions, STOP. Do not guess field names.

Fill Script

Write output/fill_YEAR.py using scripts/fill_forms.py:

add_suffix(d) — appends [0] to text field keys. Required for IRS forms.
fill_irs_pdf(in, out, fields, checkboxes, radio_values) — IRS forms. radio_values for filing status, yes/no, checking/savings.
fill_pdf(in, out, fields, checkboxes) — CA forms. Matches by /Parent chain + /AP/N keys.

Output filled PDFs to output/.

Step 9: Verify

python scripts/verify_filled.py output/f1040_filled.pdf work/expected_f1040.json

Fix any failures, re-run fill script.

Step 10: Present Results

Show a summary table, verification checklist, capital loss carryover (if any), then:

Sign your returns — unsigned returns are rejected
Payment instructions (if owed) — IRS Direct Pay, FTB Web Pay, deadline April 15
Direct deposit — recommend it for refunds; ask for bank info if not provided
Filing options — e-file (Free File, CalFile) or mailing addresses

Key Gotchas

Context

NEVER use Read tool on PDFs — use pdfplumber
NEVER read same document twice — save to work/tax_data.txt
Field discovery once per form with --compact — no --json (wastes context), no repeated --search

Field Discovery

Field names change between years — always discover fresh
XFA template is in /AcroForm → /XFA array, NOT from brute-force xref scanning
Do NOT use xml.etree for XFA — use regex (IRS XML has broken namespaces)

PDF Filling

Remove XFA from AcroForm, set NeedAppearances=True, use auto_regenerate=False
Checkboxes: set both /V and /AS to /1 or /Off
IRS fields need [0] suffix — use add_suffix()
IRS checkboxes match by /T directly; radio groups match by /AP/N key via radio_values

Form-Specific

1040: First few fields (f1_01-f1_03) are fiscal year headers, not name fields. SSN = 9 digits, no dashes. Digital assets = crypto only, not stocks.
8949: Box A/B/C checkboxes are 3-way radio buttons. Totals at high field numbers (e.g. f1_115-f1_119), not after last data row. Schedule D lines 1b/8b (from 8949), not 1a/8a.
Schedule D: Some fields have _RO suffix (read-only) — skip those.
CA 540: Field names are 540-PPNN (page+sequence, NOT line numbers). Checkboxes end with " CB", radio buttons use named AP keys.
Downloads: Prior-year IRS = irs.gov/pub/irs-prior/, current = irs.gov/pub/irs-pdf/

tax-filing