Ai Inbox Attachment Finder

SKILL.md

---
name: AI Inbox Attachment Finder
description: Builds a privacy-preserving search plan and found-file log for locating a specific email or chat attachment using metadata, likely senders, dates, keywords, and user-confirmed matches only.
version: "1.0.0"
type: prompt-flow
tags:
  - ai-productivity
  - inbox-search
  - attachments
  - admin-workflow
  - metadata
  - privacy
author: Bell (design)
---

# AI Inbox Attachment Finder

## Overview

AI Inbox Attachment Finder helps a user locate a specific form, receipt, school notice, itinerary, signed PDF, spreadsheet, image, or other attachment in email or chat without exposing sensitive contents. It produces a prioritized search plan with exact query strings, likely sender and date filters, attachment type guesses, and a found-file log for user-confirmed matches.

This skill does not read inboxes, access accounts, download attachments, expose sensitive content, or decide that a file is correct without user confirmation. It works from user-provided metadata and gives search tactics the user can run in their own mail or chat app.

## When to Use

Use this skill when the user asks about:

- Finding a specific attachment fast before a deadline
- Searching email, Gmail, Outlook, Apple Mail, Slack, Teams, WhatsApp, Discord, or another message app
- Locating a receipt, invoice, form, ticket, itinerary, school notice, signed document, PDF, spreadsheet, image, or scan
- Turning vague memory into specific search queries
- Logging candidate files without revealing private contents

**Trigger phrases:** "find the attachment", "I lost the PDF in my inbox", "search my email for a receipt", "help me find that school form", "make queries for an attachment"

## Required Inputs

Ask for metadata, not private content:

- Target artifact type: receipt, form, itinerary, signed PDF, notice, invoice, scan, image, spreadsheet, or other
- Likely sender, organization, domain, or person if known
- Approximate date range or event date
- Likely subject words, project names, merchant names, trip names, school names, or confirmation codes if safe to share
- File type guess: PDF, DOCX, XLSX, CSV, JPG, PNG, ZIP, or unknown
- Platform or app the user will search
- Deadline and what they need to do after finding it
- Any sensitive categories to avoid exposing in the chat

If the user cannot share details, use placeholders and teach the pattern.

## Workflow

### Step 1 - Define the Target File

Turn the user's memory into a short target card:

- What file is needed
- Why it is needed
- Likely date window
- Likely sender or channel
- Likely file type
- Confidence level
- Sensitive fields the assistant should not ask for

Keep the target card metadata-only. Do not ask the user to paste private document contents.

### Step 2 - Build Search Clue Buckets

Create clue buckets the user can combine:

- **Sender clues:** person, organization, domain, vendor, school, airline, bank, HR, agency
- **Date clues:** month, event date, purchase date, appointment date, application deadline
- **Subject clues:** form, receipt, invoice, itinerary, ticket, signed, completed, attached, scan, confirmation
- **File clues:** pdf, docx, xlsx, csv, jpg, png, zip, attachment
- **Amount or code clues:** only if the user says it is safe to use
- **Channel clues:** email, chat thread, shared drive notification, customer portal notification

### Step 3 - Generate Platform-Specific Queries

Produce exact query strings for the user's platform when known. If the platform is unknown, provide generic query patterns.

Examples of safe patterns:

- `from:(sender@example.com) has:attachment filename:pdf after:YYYY/MM/DD before:YYYY/MM/DD`
- `from:organization has:attachment (receipt OR invoice OR order)`
- `filename:pdf "school form" after:YYYY/MM/DD`
- `has:attachment "itinerary" "city or trip label"`
- `larger:1M filename:pdf from:domain.com`
- `"attached" "signed" "project label"`

Use placeholders when a term could reveal sensitive information. Do not fabricate exact senders, dates, or codes.

### Step 4 - Prioritize the Search Order

Create a short sequence:

1. Narrow sender plus attachment filter
2. Date window plus file type
3. Subject or body keywords
4. Organization or domain search
5. Alternate terms and misspellings
6. Trash, archive, spam, all mail, and chat files if safe
7. Related notification emails that point to a portal or drive

Explain when to broaden or narrow.

### Step 5 - Create a Candidate Found-File Log

Give the user a metadata-only log template:

- Candidate number
- Platform or mailbox
- Sender or channel
- Date
- Subject or thread label, redacted if needed
- File name or safe description
- File type
- Why it might match
- User-confirmed match: yes, no, or unsure
- Next action

The assistant should not mark a final match unless the user confirms it.

### Step 6 - Plan Safe Handling After the File Is Found

Depending on the task, suggest next actions without handling the file directly:

- Download to a named local folder
- Rename a copy with a non-sensitive label
- Upload to the official destination after user review
- Print or forward only after user confirms recipient and content
- Store a backup if allowed
- Delete duplicate downloads if they are sensitive and no longer needed

For external sending, form submission, or payment, require user review and confirmation.

### Step 7 - Produce the Attachment Finder Plan

Deliver a concise artifact with these sections:

1. **Target file card**
2. **Search clue buckets**
3. **Prioritized query list**
4. **Search order**
5. **Candidate found-file log**
6. **Confirmation rule**
7. **Next actions after finding it**
8. **Privacy limits**

## Output Format

Use this structure:

- **AI Inbox Attachment Finder Plan**
- **Target File Card:**
- **Privacy Boundaries:**
- **Search Clue Buckets:**
- **Prioritized Queries:**
- **Search Order:**
- **Candidate Found-File Log:**
- **User Confirmation Rule:**
- **Next Actions:**
- **Limits and Safety Note:**

## Safety and Boundaries

- Do not access inboxes, accounts, chats, shared drives, portals, or attachments.
- Do not ask for passwords, one-time codes, recovery codes, cookies, or account secrets.
- Do not expose sensitive document content in the chat.
- Do not infer or claim that a candidate file is correct without user confirmation.
- Do not automate forwarding, uploading, deleting, paying, signing, or submitting.
- Do not create broad searches that unnecessarily surface private or unrelated data.
- Use metadata and placeholders whenever possible.
- Keep candidate logs minimal and redacted when the file is sensitive.

## Example Prompts

- "I need to find my insurance enrollment PDF from last month. Help me build search queries."
- "Find the signed contract attachment my client sent in February."
- "Help me search my email for a school permission slip PDF from last week."

## Quality Bar

A strong response should:

- Turn vague memory into specific search tactics
- Produce exact query strings or platform-neutral patterns
- Keep all work metadata-only unless the user explicitly shares more
- Include likely senders, dates, keywords, file types, and fallback locations
- Provide a found-file log that requires user-confirmed matches
- Include safe next actions after the user finds the attachment