What prompt gets a useful AI pre-grade from card photos?
Paste this into Claude, ChatGPT, Gemini or Whizi, attach the front and back photos, and fill the brackets. It asks for ranges and doubts instead of one confident number.

Here are the front and back of a [year, set and card name] card, photographed flat and unsleeved. Act as a careful pre-grader for [PSA, Beckett or CGC]. For centering, corners, edges and surface, give a low and a high estimate on that company's scale and the one thing you saw that set each number. Estimate the centering as left to right and top to bottom percentages for the front and the back. List anything you could not judge from these photos, such as glare, a dark border or a soft focus. Do not name a value or a price for the card.
The last line matters. A chat model has no price data it can verify, and a made up number is worse than none. The doubts line matters more. It makes the model say that a bright patch might be a scratch instead of calling the surface a 7.
Is there a free AI card grader?
Yes, several, but "free" usually means one grade or a trial, and an account. The table lists what each tool says on its own page as of November 2026. A cell that reads "not stated" means the page we read did not say.
| Tool | Free use | Sign-up | Paid price | Best at, and what it says about official grades |
|---|---|---|---|---|
| CardGrade.io | One full AI grade free for eligible new accounts | Free signup | Plans from $4.99 a month | A set workflow with 16 inspection zones and a Chrome extension. Says the result is an image based estimate and actual grades may differ |
| CardGrader.AI | "Try Free" on its button; limit not stated | Sign in; steps not stated | Shown in the App Store and Google Play | Phone scanning. Claims 99.2% accuracy itself; no disclaimer about PSA, BGS or CGC found |
| PSA app (Google Play) | Not stated on the listing | Not stated | Not stated | Tracking a PSA submission and collection. The listing mentions neither scanning nor AI |
| Ximilar | Grader open to users of all plans for testing | Account in its app | Not stated in the article | Developers. Scores centering, corners, edges and surface; says it is not a replacement for PSA, BGS, SGC or CGC |
| ChatGPT | 3 file uploads a day on Free | Not stated | Not stated | A general model that reads photos and follows your rubric |
| Claude | 20 files per chat | Not stated | Not stated | Long, careful written comparisons across many photos |
| Whizi | 5 photo uploads a day on the free plan | Account needed | Pro $29.99 a month | One chat for front, back, your rubric and a second model |
Whizi loses on three counts. CardGrade.io has a fixed workflow that measures 16 zones and runs inside eBay and TCGPlayer listings through its Chrome extension, and a chat doesn't have that. Ximilar scores centering on the Beckett scale as a number, where a chat model estimates it by eye. And the PSA app is the one that tracks your real submission.
The sources differ in kind. CardGrade.io's guide to AI card grading lists its price and its limits. CardGrader.AI's blog post names no price. It calls its own result "industry-leading 99.2% accuracy", and the same post says AI grading systems in general reach "95-99% accuracy" against PSA, BGS and SGC. Both are the vendor's own claims, with no method shown. Ximilar's article is the most candid, and it comes from a company that sells the technology.
Is an AI card grader accurate?
No tool we read publishes an accuracy test against PSA, BGS or CGC grades that we could verify. CardGrader.AI claims 99.2% accuracy in its own blog post, and that is the only accuracy number any of them states for itself. Treat it as marketing until someone shows the method.
The vendors that are careful say as much. CardGrade.io says its result "does not guarantee eligibility, value, a sale price, or a profitable submission." Ximilar says AI "can make mistakes," that photo quality drives the result, and in one of its own examples that the system "misses the surface by quite a large margin."
Surface is the hard one. A flat photo can't show a faint print line or a shallow dent that a grader sees with the card tilted under a light. Centering is the easiest, because it is a measurement the camera can take. Use an AI pre-grade to catch obvious problems before you pay a company. Don't use it to decide that a card is a 10.
What does PSA say about how its grades are set?
Each company grades on its own scale with its own words, and an AI estimate has to pick one. PSA grades from 10 down to 1. Its published names run Gem-Mt 10 (Gem Mint), Mint 9, NM-MT 8 (Near Mint-Mint), NM 7, EX-MT 6, EX 5, VG-EX 4, VG 3, Good 2 and PR-FR 1 (Poor to Fair), per its Grading Your Cards article. That article is undated, and it puts Gem Mint centering at approximately 55/45 to 60/40 on the front and 75/25 on the reverse.
The PSA grading standards page is the current one, and it gives no centering table. Its one ratio is a 70/30 example, used to say that a card can measure for a grade and still be held back on eye appeal. It does not mention AI. So treat any centering figure as that older article's, and a tool that returns a hard cutoff as claiming more precision than PSA does.
Beckett and CGC publish their own scales at Beckett and CGC. Both sites blocked our reader, so this page states nothing about either scale. Open the company page for the one you plan to submit to before you set it as the target in a prompt.
On AI, the PSA standards page says nothing. That is an absence in one page, not proof about how a company works inside. If a tool tells you it is "PSA standard," it means it is trained or prompted to imitate a published scale. PSA did not certify it.
Can ChatGPT or Claude grade my cards?
They can pre-grade from photos, and the prompt above is how. ChatGPT Free allows 3 file uploads a day per OpenAI's upload FAQ, so a front and back uses 2 and a second card has to wait. Claude takes 20 files per chat per its upload page, which fits a batch. AI that can analyze images lists the limits for Gemini and Copilot too.
A chat model is not trained on a grading company's data, and it won't measure your centering in pixels. It reads the photo the way a person glancing at it would. That is fine for spotting a bent corner or a miscut. It's weak for a close call between a 9 and a 10.
What does a good AI pre-grade look like on a real card?
A good pre-grade names a measurement or a flaw for every number and says what it could not see. Here are three cards a collector might send, with the prompt line to add and what a useful reply must contain. These replies are described for this page, not output from a named model.
- A modern holo, front and back, bright and sharp. Add: "This card is a holo. Tell me which bright areas could be glare and which look like real scratches." A good reply gives front centering as two percentages, marks the holo surface as a doubt, and says the back should be judged on its own.
- A vintage card with soft corners and a dark border. Add: "Centering is hard to see on this border, so say how sure you are." A good reply says the dark border makes the centering estimate rough, rates each corner separately, and refuses to name a grade for the surface from one shot.
- A card with a crease you can see in the light. Add: "I can see a crease near the top left in person but it is faint in the photo. What would a grader notice here?" A good reply says a crease usually caps the grade by a wide margin and asks for a second photo at an angle.
If the reply to any of these gives one grade and no doubts, send it back with "give a range and list what you could not judge."
What AI card grading prompts work first try?
These prompts cover the steps around the grade: the photo, the estimate, the comparison and the listing. Fill the brackets and attach the photos each one names.
Before you shoot
- Photo check: "Here is my photo of the front of [card]. Tell me if the light, angle or crop would stop you grading centering, and what to change."
- Glare finder: "Mark every bright patch in this photo that could be glare, as top left, center and so on, so I can retake it."
The estimate
- Full pre-grade: use the prompt at the top of this page.
- Centering only: "Estimate the centering of this [card] as left to right and top to bottom percentages, and say which border you measured and how sure you are."
- Corners one by one: "Describe each of the four corners in these photos as sharp, slightly soft or worn, and say which are hardest to judge."
- Surface doubts: "List every mark on the surface that could be a scratch, print line or dust, and say which a second photo at an angle would settle."
Against the scale
- Scale check: "Using [PSA, Beckett or CGC] published scale, what is the highest grade this card could reach if the surface is clean, and which of my four numbers holds it back?"
- Before and after: "Compare my photos from [date] and today of the same card and tell me if anything changed, such as new wear or a bend."
The decision
- Send or skip: "Given these four estimates, list the reasons I might send this card to [company] and the reasons I might hold it. Do not mention what the card is worth."
- Batch triage: "Here are [number] cards. Rank them from the most to the least likely to reach [grade], with one reason each, and flag any photo you could not judge."
- Seller description: "Write a plain description of the condition of this card from these photos for a listing, naming every flaw you can see and none you cannot, and say it is ungraded."
- Label check: "Here is a photo of a graded slab. Read the grade, the company and the certificate number, and tell me which part is hard to read. Do not tell me if it is genuine."
What mistakes ruin an AI card grade?
Five mistakes cause most bad pre-grades, and each has a quick fix.
- Photographing the card in its sleeve. The fix: take it out, and keep the slab or loader for the last photo only.
- Reading glare as damage. The fix: take two photos with the light moved and compare.
- Sending only the front. The fix: send both sides, because a bad back corner or a stain is half the grade.
- Treating one number as a grade. The fix: ask for a range and plan around the low end.
- Asking what the card is worth. The fix: leave value out. A model cannot verify sale prices, so it will guess.
How do you use Whizi as one app instead of a pile of single purpose tools?
Whizi is one chat where you attach the front and back photos, paste your own rubric and read the answer, instead of juggling a scanner app, a spreadsheet and a chatbot. An uploaded image can be attached to a chat message for the model to look at, and the free plan allows 5 photo uploads a day alongside 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card.
Pro carries Claude, GPT and Gemini in one app, and its first week is $1.99. Compare runs two models on one prompt side by side on Pro and Powerhouse. Send the same photos and the same prompt to two models and read only where their centering and surface notes disagree. Web search is on by default in a chat, so the model can look up a grading company's own published scale. See Whizi pricing for what each plan holds.
Whizi won't replace a dedicated grader for a batch. It doesn't have a measuring tool for centering, and CardGrade.io is quicker for a card that you want run through a set workflow.
When is an AI pre-grade the wrong tool?
It is the wrong tool for any decision that needs a certified grade. A buyer who wants a graded card wants the slab, and no AI estimate sits in one. It is also wrong for authenticity. A model can't tell you whether a card is real, and CardGrade.io says it is independent and not affiliated with the grading companies. Use the company's own verification on the certificate number for that, and how to fact check AI answers for the habit of opening the source.
A pre-grade is also weak on very old cards, on dark borders and on foil that throws glare. In each case the doubts line in the prompt is doing the work, so keep it in.
- Photograph front and back, unsleeved, flat and in soft light
- Retake any photo with a bright patch that could be glare
- Ask for a low and high estimate for centering, corners, edges and surface
- Name the grading company and say which scale to use
- Inspect every flaw the model named with a loupe before you decide
- Leave card value out of the prompt
Frequently asked questions
Is there a free AI card grader?
Yes. CardGrade.io states one full AI grade free for eligible new accounts, as of November 2026, and CardGrader.AI's button says "Try Free" without stating a limit. Whizi's free plan and ChatGPT Free also read photos, within their daily upload limits.
Can AI replace a PSA grade?
No. An AI estimate is a prediction from photos, and it carries no company's label. CardGrade.io calls its result "Predicted condition, not an official third-party grade," and Ximilar says its intention is by no means to replace PSA, BGS, SGC or CGC.
How accurate is AI card grading?
No tool we read publishes a verified test against PSA, BGS or CGC grades. CardGrader.AI claims 99.2% accuracy in its own blog post, a vendor claim with no method shown, and Ximilar shows an example where its model misses the surface by a large margin. Treat an estimate as a range.