Invoice Chaser

Invoice Chase

Quick start

Ask the user for their AR aging data, score each customer by payment history, draft a tone-matched reminder for each overdue invoice, and present them to the owner. Nothing sends until the owner says so.

User: "who owes me money"
→ Ask the user to upload or paste their AR aging report (e.g., a QuickBooks CSV export)
→ Cross-reference any payment records they also provide (e.g., a PayPal settlements export)
→ Score each customer: good-payer / occasionally-late / repeat-late
→ Draft tone-matched reminders
→ Show summary table + drafts. Wait for "send these."

Setup (first run only)

Ask the owner two questions before running for the first time:

1. Mail preference: "Do you use Gmail or Apple Mail for drafts?" — store the answer; use it when describing where to queue non-PayPal drafts.
2. Stripe: "Do you also invoice through Stripe? If so, export your Stripe overdue invoices and I can include them in the sweep." — if the user provides a Stripe export, fold those overdue invoices in alongside the QuickBooks data.

Do not ask again on subsequent runs.

Workflow

1. Get overdue receivables. Ask the user to upload or paste their AR aging report (e.g., a QuickBooks CSV export) covering all invoices more than 1 day past due. If the user also provides a Stripe overdue-invoices export, include those invoices too.

2. Cross-reference payment history. If the user also provides a payment record (e.g., a PayPal settlements export), use it to catch invoices that may already be paid. When asking for it, request:

   • Settled transactions only (filter out pending and denied transactions that inflate the list)
   • A recent window — the last 7 days ending today is usually enough to catch payments not yet reflected in the AR aging

   If the user hasn't provided any payment record: skip the payment cross-reference for this run. Flag all customers in the batch as "payments not cross-checked — verify manually" in the summary table. Proceed to scoring using the AR aging history only. Do not silently drop the caveat.

   If a customer shows a settled payment within the provided window, flag as "possibly paid — verify" and exclude from the draft queue.

3. Score each customer. Read reference/tone-matching.md for scoring logic. Result: good-payer, occasionally-late, or repeat-late.

4. Draft reminder emails. One email per customer — consolidate multiple overdue invoices into one email. Match tone to score. See reference/examples/gentle-reminder.md and reference/examples/firm-reminder.md.

5. Present drafts to owner. Show a summary table first:

   Customer	Amount Due	Days Late	Tone	Send via
   Acme Corp	$1,200	18 days	Gentle	PayPal
   Smith LLC	$450	47 days	Firm	Gmail draft

   Then show each draft email in full. Wait for owner to say "send these" or approve individually.

6. Send or queue — only after approval.

   • PayPal invoices: send the reminder via PayPal.
   • Non-PayPal invoices: queue as a draft in the owner's configured mail app.
   • Never send without explicit approval.

7. Report what happened. List what was sent, what was queued as draft, and what was flagged (possibly paid, excluded).

Approval gates

• Never send or queue a draft without explicit owner approval. Present all drafts first; wait for the go-ahead.
• Never include a customer who paid in the last 14 days (per any payment record the user provided). Flag as "possibly paid — verify" instead.
• Never send to a customer not in the AR aging the user provided (or the Stripe export, if they provided one). No reminders from memory alone.
• One approval covers one batch. Adding a customer or changing a draft after approval starts a new round.

Reference

• reference/tone-matching.md — scoring logic, tone guidelines, subject line formulas
• reference/gotchas.md — known failure modes
• reference/examples/gentle-reminder.md — good-payer email example
• reference/examples/firm-reminder.md — repeat-late-payer email example

Zeplik output presentation

Present the final deliverable as a single polished artifact: clear headings, tables where the content is tabular, fenced code where it is code. Lead with the deliverable itself; keep process commentary to a single short line. If the skill produced multiple files or sections, end with a compact list of them with one-line purposes.