--- name: "before_after_proof_pack" description: "Turn real before and after results into proof a buyer can trust: a capture protocol that keeps lighting, angle, crop and scale comparable, a claim check that traces every caption claim to source records, honest captions that state what changed and what did not, and a pack format for listings, posts and case pages. Use when building, fixing or auditing a before and after pair, gallery, slider, transformation post or case page for a product, a service, a place or a person." --- # Before and After Proof Pack ## Purpose Most before and after pairs fail before anyone writes a word. The light moved, the camera moved, the after got a filter, the best case got picked and shown as the normal case, and the caption claims a number nobody recorded. A tool that tidies two supplied screenshots into a neat pack cannot fix any of that. This skill works earlier and later than the tidy up: a capture protocol that makes a pair honestly comparable in the first place, a claim check against the records behind the images, and a pack format that carries the evidence with the pictures into a listing, a post or a case page. ## Workflow 1. **Start before the before.** If the work has not started, run the capture protocol in `references/capture-protocol.md` first. Set one position, one lighting setup, one angle and one landmark, give each an ID, write the IDs and the distance down, and take the before before tools come out or the product is opened. A before taken late cannot be retaken. 2. **Build the manifest.** One pair, one manifest JSON, following the examples in `assets/manifest-service.example.json` and `assets/manifest-listing.example.json`. Record subject, dates, setup IDs for both images, edits, source records, measurements with their sources, the caption claims, what changed, what did not change, what the photos cannot show, consent, material connection and typicality. 3. **Run the checker.** Run `python3 bin/proof_pack.py check `. It fails the pack on mismatched setup IDs, distance drift over 10 percent, after only edits, missing consent for a person or a private place, numbers in claims that no record holds, a time claim that disagrees with the dates, an undisclosed connection, and an atypical result with no generally expected result stated. Fix the record or cut the claim for every fail. Treat warnings as edits to make, not as decoration. 4. **Run the claim check by hand.** Follow `references/claim-check.md`. List the implied claims the images make, not just the written ones. Check the cause: name anything else that changed between the dates. Check typicality and selection: how this case was picked, and what the usual result is, with its basis. Apply the harder health and body checks when the pair touches health, weight or skin. 5. **Build the pack.** Run `python3 bin/proof_pack.py build --out pack.md`, then finish the captions in plain language using `assets/caption-templates.md`. Format the pack for its surface using `references/pack-formats.md`: short at a listing, disclosure first in a post, full record on a case page. All three versions carry the same claims. 6. **Check the surface before it ships.** Confirm consent covers this surface, identifiers are out of frame, the pair still reads honestly at phone size and after platform compression, and, for paid use, the current platform policy for the exact ad format, checked live that day with the date noted in the pack. For health and body ads on Meta, that check is mandatory, because the standard changes in place. 7. **File the proof.** Keep the manifest, the two original unedited files, the consent record and the pack together. If the originals are lost, retire the pack. The pack is the claim, and a claim with no file behind it does not run. ## Output Contract - A manifest JSON for the pair that passes `bin/proof_pack.py check` with zero fails, or a plain statement of which fails could not be fixed and what was cut because of them. - A pack with these sections in order: header with dates, interval and setup, the dated pair, what changed, what did not change, what the photos cannot show, source records, disclosures, and the caption versions the surface needs. - Every number in any caption traces to a source record named in the pack. Every result claim states its interval. Every atypical result states the generally expected result next to the pair. - Two worked packs ship with this skill as the standard to copy: a service result pack and a product listing pack, in `references/pack-formats.md` with their manifests in `assets/`. Their data is marked EXAMPLE, because they teach the shape of a pack and are not real customer results. ## Operating Rules 1. Same subject, same setup, or no comparison claim. If position, lighting, angle or landmark differed and the difference was not part of the job itself, the pair can show that work happened and nothing more. Rewrite the claim down to what the weaker pair supports. 2. No edit applied to the after image only, ever. No filter, smoothing, reshaping, object removal or color shift on one side of a proof pair. A disclaimer does not repair a misleading image. 3. Never invent a number, a date, an interval, a rating or a result. Untraced means cut. The checker enforces this for listed claims, and the rule covers text the checker cannot see. 4. Results not typical is not a typicality statement. Under the FTC Endorsement Guides revised in 2023, state the generally expected performance in the depicted circumstances, matched in scope and time to the claim, or show the basis that makes the result representative. Facts and sources: `references/verified-facts.md`. 5. Disclose every material connection in the caption, next to the claim, in plain words: paid, free product or service, employee, family, affiliate, prize chance. If the surface cannot carry the disclosure clearly, do not run the pair there. 6. Fake proof is out completely, not a gray area. No stock pairs, no other firm work, no model who never used the product, no generated images presented as a customer result. The FTC Consumer Review Rule, 16 CFR Part 465, effective October 21, 2024, bans testimonials that misrepresent whether the person used the product or what their experience was. Generated or stock material may run only as a labeled illustration with no result claim. 7. Health and body claims need evidence that one happy result is not. Follow the health section in `references/claim-check.md` and cut condition, cure and timeframe promise language that the records do not support. 8. Get signed, dated consent with a stated scope before publishing a person or the inside of a private place, and keep faces, addresses and documents out of frame unless consent covers them and the pack needs them. 9. State what did not change and what the photos cannot show in every pack. A pack with only a changed list is an ad. A pack with all three lists is proof. 10. Write plain. No hype words, no miracle language, no pressure lines about the viewer body or home. Describe the work and the record, and let the pair do the persuading.