Save the Info in a Screenshot Before Deleting It | SnapSweep

The short answer

You do not have to choose between the space a screenshot takes and the information inside it. SnapSweep's Info Vault reads the screenshot while it still exists, lifts out the things worth keeping, and then lets you delete the image.

The extraction covers phone numbers, URLs, addresses, tracking numbers and verification codes — plus email addresses and WeChat IDs — and each one arrives with a button that performs the matching action: call, open, navigate, send, search, or copy.

SnapSweep runs all of it on device with no account and no cloud, and deletions still go through iOS's native trash behind a 30-day recycle bin. SnapSweep is on the App Store — but the rest of this page is about how the extraction decides what is worth keeping, because that is where this kind of feature usually falls apart.

Why the screenshot never gets deleted

A screenshot pile survives not out of sentiment but because a screenshot is frequently the only copy of something practical: a courier's tracking number, the address of an office you visit tomorrow, a code you will need in four minutes. Delete the image and the information goes with it. Every cleaner offering only "delete or keep" is quietly asking you to gamble, and you decline every time.

There is a second reason, and it is worse. The information is often already useless and you cannot tell. A screenshot from two years ago holds a tracking number for a parcel that arrived long ago and a discount code for a shop you no longer use — and nothing about looking at it tells you which is which.

Info Vault collapses that uncertainty before the delete, not after it.

The moment the copy is taken

Extraction is not a separate step you have to remember; it is attached to the swipe gesture that does the deleting.

• Swipe left to delete, and the app runs text recognition on the screenshot and saves whatever it finds.
• Swipe up to favorite, and the same pass runs — the photo is staying, but the information is banked anyway.
• Swipe right to keep, and nothing is extracted. You kept it, so there was nothing to preserve.

The recognition itself is unglamorous on purpose. SnapSweep asks the photo library for a fresh analysis image with its longest edge set to 1400 points, not the card thumbnail you were just looking at. That matters more than it sounds: a review card crops a long screenshot down to a phone-shaped window, and what remains is often too coarse to read a 15-digit tracking number digit by digit. The pass runs with Simplified Chinese and English recognition, covering both a domestic courier app and an overseas order confirmation.

What counts as an item, and what gets thrown away

The extraction is not "find anything that looks like text". Each candidate is scored, and only the ones that clear a bar are stored.

Every score is the OCR engine's own confidence multiplied by a strength assigned to the rule that found it, and the floor is 0.5. Below it the candidate is dropped rather than stored — the deliberate trade is to miss some real items rather than fill the vault with junk you would have to clean up yourself.

The rule strengths explain what the extractor is actually confident about:

Read that as arithmetic. A phone number in a crisp line passes easily; an address at strength 0.7 needs a sharp read, and a tracking number backed only by a carrier word needs OCR confidence above roughly 0.83. Weaker evidence demands clearer pixels.

Why a random number is not a tracking number

Tracking numbers are the case where a naive extractor is most obviously wrong, because a screenshot of any order contains several 10-plus-digit numbers that are not tracking numbers.

The rule refuses to treat a bare number as one. A candidate must be 8 to 24 alphanumeric characters, and if purely numeric it must also be 10 to 22 digits long and backed by a carrier word — UPS, FedEx, DHL, USPS, the Chinese carriers, or the generic tracking, waybill, shipment — on the same line or a vertically adjacent one. Adjacency is measured: the neighbour has to sit within 0.15 of the vertical position to count.

Two shortcuts are recognised outright, because the format is unambiguous: a UPS number beginning 1Z and an SF Express number beginning SF. A number passing the mod-10 check digit GS1 and USPS IMpb share is promoted to a stronger score than a keyword match alone.

The same discipline governs verification codes. A four-to-eight digit string is a candidate only when a keyword — verification, verify, passcode, OTP, security code, code, or 验证码、校验码、 动态码、一次性密码 — introduces it, with the digits after the keyword, on the same line or an adjacent one. Without that keyword the extractor only tries when the screenshot is already classified into the verification-code category, and it discards four-digit values between 1900 and 2100. Searching a whole screenshot for digit runs would sweep up amounts, floor numbers and years; position is the actual evidence.

Every button does what it says

The other half is that a saved item is usable, and this is where vaults quietly lie: a row labelled "Navigate" that copies to the clipboard wastes the one tap you had. SnapSweep resolves the action from the item's kind and derives the label from that action, so the two cannot disagree:

• Link — opened directly, but only if it parses to http or https. Anything else becomes Copy; treating arbitrary text as a link produces a silent no-op rather than an honest failure.
• Phone — Call, through a tel: number built from the digits alone, keeping a leading + if present.
• Email — Send, via mailto:. Address — Navigate, through an Apple Maps query.
• Tracking number — Search, but only for the five carriers with a real tracking page: UPS, FedEx, USPS, DHL and SF Express. For any other carrier, or none recognised, the button says Copy rather than assembling a search-engine query for a number it cannot vouch for.
• Verification code, WeChat ID, anything else — Copy.

WeChat IDs are only captured when the label 微信 or WeChat directly introduces the account name; taking the first alphanumeric run from any screenshot that merely mentions the word returns an ordinary English word.

The vault does not accumulate duplicates, and it forgets on purpose

Two behaviours make this usable over months.

It deduplicates on identity, not on appearance. An item's identity is its kind plus a normalized value: URLs lowercased with trailing slashes stripped, phone numbers and codes reduced to digits, tracking numbers uppercased with spaces and dashes removed, addresses whitespace-collapsed. Save the same tracking number from three screenshots and you get one entry, refreshed — not three rows to reconcile.

It refuses to extend a code's life. Re-extracting the same screenshot merges into the existing entry and keeps the original expiry — correct behaviour, and easy to get wrong: a code is valid because of when it was issued, not because you scrolled past its screenshot for the fourth time.

Codes carry the only expiry in the vault — ten minutes from extraction — and an item is flagged Expiring in the final 24 hours, with that same window driving the "Expiring soon" counter so badge and number cannot disagree. Expired items are filtered out of every read, purged on load, and removable on demand. A row shows its context — the carrier for a tracking number — with a relative timestamp beside it. You can filter to All, URLs, Contacts, Packages or Addresses, and search across value and context.

The whole thing is one JSON file in the app's Documents directory, written with complete-until-first-authentication file protection, on device, with no account.

What SnapSweep does differently

Most screenshot cleaners treat the image as the product: find the duplicates, hand them over, let you swipe. SnapSweep does that too — near-duplicate grouping and a video size analyzer with HEVC suggestions are both in the box. What it also does is treat the content as something to preserve before the image goes.

The difference is where that happens. Extraction rides the delete gesture rather than living on a separate screen you must remember to visit, it runs against a full-resolution image rather than the crop you were reviewing, and the scoring keeps the vault small. Most importantly it is lossy in the right direction: an item it cannot vouch for is not stored, and a code it cannot date is not kept alive.

Is it worth it

If your library is mostly screenshots you have never opened twice, none of this matters; classification and batch delete already cover it.

If your library is full of the other kind, the question is not whether deleting is safe but whether you can tell which ones are. Answer that first, and the delete stops being a gamble.

Download SnapSweep on the App Store, or read how the review gesture works in the bulk-delete guide first.