agentleFS
geodgeod/fsdmoney/.claude/skills/income/SKILL.md

income

Skill0 starsChanged 6 months agoMITView on GitHub

Interview-driven workflow for capturing and analyzing W-2 income. User either drops one or more W-2 PDFs or walks through the boxes interactively, and gets back a short descriptive breakdown — gross, tax stack (federal, state, FICA), pre-tax deductions, take-home — plus internal consistency checks on the W-2. Strictly descriptive — describes what the W-2 says, never recommends withholding changes, contribution rates, or comp moves. Use whenever the user wants to understand their W-2 income at a glance. Trigger on phrases like "what do I earn", "analyze my W-2", "income breakdown", "take-home pay", "effective tax rate", "income analysis", or when the user drops a W-2 PDF. Future versions will add 1099, K-1, investment income, rental, and retirement income — currently W-2 only.

The file

We split this skill into the rules it gives an agent

9 rules. About 0 are common, 6 are distinctive, and 3 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈6

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 3

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    List first: If the user points at a folder, ls it and enumerate every file with format + size.

    line 47
  2. Very specific:

    Image-only PDFs: If extract_text() returns empty or near-empty, the PDF is a scan.

    line 49
  3. Distinctive:

    JPG/PNG/HEIC directly: Read as image.

    line 50
  4. Distinctive:

    Every Box 12 entry (code + amount) — *including* C (GTL imputed), D (401k), Y (NQDC §409A), DD (employer health), and any others present

    line 56
  5. Distinctive:

    If you can't extract cleanly, fall back to Path B for any uncaptured boxes — don't guess.

    line 61
  6. Very specific:

    Ask for these boxes per W-2.

    line 67
  7. Distinctive:

    If a check fails, surface it to the user before proceeding: *"This W-2 shows X, but the IRS formula says Y.

    line 97
  8. Distinctive:

    Do NOT subtract Box 12 DD (employer health coverage) — informational only.

    line 127
  9. Distinctive:

    If the user has more than one W-2 (mid-year job change, both partners working, or both):

    line 133

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Income

What this skill does

Walks the user through their W-2 income. They can either drop one or more W-2 PDFs (the skill extracts every box) or walk through the boxes interactively. Output is a short descriptive analysis shown inline in the conversation — no separate report file is written.

v1 scope

W-2 employment income only. Other income types (1099, K-1, investment, rental, retirement, other) are not yet supported. If the user mentions one, acknowledge it'll come later and proceed with the W-2 portion.

If the user explicitly insists on adding a non-W-2 source mid-conversation (e.g., drops a rental income spreadsheet and says "add this"), do a descriptive-only ingest rather than refusing twice: read the file, surface a cash view and a tax view, flag every missing input (depreciation schedule, cost basis, etc.), and roll it into the household total as a clearly-labeled out-of-scope addition. The same banned-vocabulary rules apply. Do not invent Schedule E categories that aren't supported by the source data.

What this skill does NOT do

This is descriptive only. Never:

  • Recommend withholding changes ("you should adjust your W-4")
  • Recommend contribution rates ("max your 401(k)", "increase your HSA")
  • Comment on comp structure, RSU/ESPP timing, or tax-bracket optimization
  • Project or model anything forward — that's downstream

Those belong in a downstream financial-planning skill that takes this skill's output as one of its inputs. If the user asks for recommendations, redirect.

Banned vocabulary in output: you should, consider, recommend, worth, would be wise, target, under-saving, over-withheld, under-withheld, optimize, improve. State facts; don't editorialize. The user (or a downstream skill) decides what's high or low.

The flow

1. Ask

"Let's look at your W-2 income. Two ways — drop your W-2 PDF(s) in a folder and tell me the path, or walk through the boxes with me (takes a couple of minutes per W-2). Which works?"

Skip the menu if the user has already signalled — e.g., they dropped a file in the conversation, or they said "I'll just tell you the numbers." Also ask up front: how many W-2s — just yours, or also a partner/spouse?

2. Capture

Path A — drop W-2 file(s):

Files may arrive as PDFs (text-based or image-only scans), JPG/PNG photos, or HEIC→PDF exports. Always parse the file at the attached path directly. Do not assume any image pasted into the chat is the contents of an attached file path — they may be different documents, and conflating them produces silently wrong analysis.

Parsing pipeline:

  1. List first. If the user points at a folder, ls it and enumerate every file with format + size. Confirm with the user which file corresponds to which person before parsing.
  2. Text extraction. Try pypdf (PdfReader(...).pages[0].extract_text()). On macOS, pdftotext is not installed by default — don't rely on it. pdfplumber is a layout-sensitive fallback.
  3. Image-only PDFs. If extract_text() returns empty or near-empty, the PDF is a scan. Extract the embedded image (page.images[0].data) to a temp file and read it as an image — most W-2 PDFs from phone scanners or HEIC exports fall into this bucket.
  4. JPG/PNG/HEIC directly. Read as image.

Extract these fields per W-2:

  • Employee name, employer name, year
  • Boxes 1, 2, 3, 4, 5, 6
  • Every Box 12 entry (code + amount) — including C (GTL imputed), D (401k), Y (NQDC §409A), DD (employer health), and any others present
  • Box 13 checkboxes (retirement plan, statutory employee, third-party sick)
  • Box 14 — capture if present (CA SDI, CA VPDI, NJ SDI, NY SDI, etc. are mandatory cash deductions and DO affect take-home)
  • Boxes 15–17 (state, state ID, state wages, state tax)

If you can't extract cleanly, fall back to Path B for any uncaptured boxes — don't guess.

Filename vs Box e mismatch. If the filename suggests person A (e.g., w2-alice.pdf) but Box e on the form names person B, surface the mismatch once and ask which to trust. Don't loop on it. The user is authoritative — once they answer, proceed and don't re-litigate.

Path B — interview entry:

Ask for these boxes per W-2. Use AskUserQuestion or accept a freeform paste:

BoxWhat to ask
1Wages, tips, other comp (the federal taxable number)
2Federal income tax withheld
3Social Security wages
4Social Security tax withheld
5Medicare wages and tips
6Medicare tax withheld
12Any codes + amounts. Common: D = 401(k) traditional, AA = Roth 401(k), W = HSA, Y = NQDC §409A current-year deferral, C = GTL > 50K (imputed non-cash, included in Box 1/3/5), DD = employer health cost (informational only)
14Ask if there's anything in Box 14. State disability lines (CA SDI, CA VPDI, NJ SDI, NY SDI, HI TDI, etc.) are mandatory and reduce take-home — capture the amount. RSU/ESPP entries here are usually informational only; if mentioned, prompt: "Is the RSU/ESPP amount already included in Box 1? Most employers include it."
17State income tax withheld (skip if no state tax)

Skip blank boxes.

3. Confirm + analyze

Show the captured values back to the user, run the internal consistency checks, and produce the inline analysis. Everything in the conversation — no files written.

Internal consistency checks (run all five; surface any failure rather than silently proceeding):

#CheckTolerance
1Box 5 ≥ Box 1exact
2(Box 5 − Box 1) reconciles to the sum of Box 12 deferral codes that reduce Box 1 but not FICA — primarily D (401k traditional), E (403b), G (457b), S (SIMPLE), Y (NQDC §409A current-year deferral). If the gap doesn't reconcile, pre-tax health/FSA likely accounts for the residual (those reduce both Box 1 and Box 5 and so are invisible here — the residual size is itself informative)small residual OK
3Box 3 capped at SS wage base for the year (168,600 in 2024; 176,100 in 2025)exact
4Box 4 = Box 3 × 6.2%1.00
5Box 6 = Box 5 × 1.45% + max(0, Box 5 − 200,000) × 0.9% (Additional Medicare for employee withholding)1.00

If a check fails, surface it to the user before proceeding: "This W-2 shows X, but the IRS formula says Y. Worth double-checking the document."

Display rule for all numeric output. Do not prefix numbers with $. The Claude Code markdown renderer interprets $<digit> patterns as math-mode tokens and silently strips the leading digits, producing wrong-looking tables (e.g., ,048,350.49 instead of 1,048,350.49). Use plain digits with thousands separators. If a currency label is needed, add it as a column header or table caption, not inline before each number.

Inline analysis sections — produce all of these as compact markdown tables in the response:

  1. Headline — Gross (Box 5), Total tax withheld (Box 2 + Box 17 + Box 4 + Box 6), Pre-tax cash deductions visible (sum per Net formula below), Net take-home computed. Each with amount and % of gross.
  2. Tax stack — Federal, State, SS, Medicare each as a row.
  3. Pre-tax / non-cash items — list every Box 12 entry present and Box 14 SDI/VPDI. For each, mark whether it (a) reduces cash take-home, (b) is informational only, or (c) is imputed non-cash income included in Box 1/3/5. Footnote: pre-tax health premiums and FSA contributions reduce Box 5 silently and are invisible on the W-2.
  4. Take-home run-rate — Annual net; monthly equivalent; bi-weekly equivalent (26 periods).
  5. Caveats — anything that didn't reconcile, any boxes the user skipped or the parser couldn't extract, and a one-liner about v1 scope (other income types not yet covered).

Net formula:

Net cash take-home
  = Box 5
  − Box 2                       (federal income tax)
  − Box 17                      (state income tax)
  − Box 4                       (Social Security)
  − Box 6                       (Medicare incl. Additional 0.9%)
  − Box 14 mandatory disability (CA SDI, CA VPDI, NJ SDI, NY SDI, HI TDI, etc.)
  − sum of Box 12 codes that reduce cash:
      D (401k traditional), E (403b), G (457b), S (SIMPLE),
      AA (Roth 401k), BB (Roth 403b), EE (Roth 457b),
      W (HSA), Y (NQDC §409A current-year deferral)
  − Box 12 C (GTL imputed)      — subtract because it is non-cash income
                                  included in Box 1/3/5 but never received as cash

Do NOT subtract Box 12 DD (employer health coverage) — informational only.

Tax character is always ordinary for W-2 wages — only mention if the user asks.

Multiple W-2s

If the user has more than one W-2 (mid-year job change, both partners working, or both):

  • Repeat the capture step per W-2
  • Roll up the analysis: headline gets per-person rows + household total; tax stack and deductions sum across W-2s
  • Surface any inconsistency between W-2s for the same person (e.g., total Box 3 across all W-2s for one person exceeds the SS wage base → excess Social Security withheld, which is a fact worth stating but without prescriptive framing)

Year alignment

All inputs combined into a single household view must share the same tax year. If the user attaches sources from different years (e.g., 2025 W-2s + 2024 rental statement), flag the mismatch and either ask for the matching year or label the combined view explicitly as "most-recent-available, mixed years" with a clear caveat. Don't silently merge across years.

What to do when finished

End the conversation cleanly. Don't push the user toward anything. If they explicitly ask "what's next?", mention that the downstream financial-planning skill (when it exists) takes income + expenses + investment-analysis as inputs.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.