Skip to content
Stitchproof

Every row, proven before you cast on.

Paste a knitting pattern. Stitchproof reads it, counts every stitch in every size, and shows you exactly where the numbers don't add up, and how to fix them.

Free, no account. Standard patterns are read entirely in your browser.

Doesn't add upHarbour Beanie · sample

Rnd 49

size Adult L

The instructions make 65 stitches; the pattern says 64.

5 later rounds go wrong because of this one.

Suggested fix

Change 8 to 7 in Rnd 49 (size Adult L only).

−Rnd 49: *K5 (6, 8), k2tog; rep from * around. (48 (56, 64) sts)
+Rnd 49: *K5 (6, 7), k2tog; rep from * around. (48 (56, 64) sts)

With this edit, every row in every size adds up.

October 22, 2026

Etsy banned AI-written patterns. Buyers still can't tell which ones they are.

Etsy now prohibits AI-generated knitting, crochet and other craft patterns, after admitting that many buyers “can't tell which items involve AI-generated content.” A knitwear designer told CBC her sales had halved since those patterns appeared.

A rule needs evidence. A pattern whose stitch counts cannot work is evidence anyone can check, and Stitchproof finds it in seconds: which row, which size, by how many stitches. Designers get the same check before they publish.

AI reads the words. Code counts the stitches.

Language models write pattern-shaped text without keeping a running total. That's why AI-generated patterns so often don't work. Stitchproof only trusts AI to translate. Every count is plain arithmetic you can check.

STS
One round of a hat crown, *K2, k2tog; rep from * once more, set as type. Each slug is as wide as the stitches it uses and carries one pin per stitch it makes; every pin the feeler climbs turns the dial one division. The marker holds the count the pattern states (here 6, then a typo'd 7), and the lamp shows whether the needle stops on it. Stitchproof's counter runs this same uses/makes table on every row of every size.
  1. 1

    Read

    The offline reader understands standard notation. Only lines it can't read are sent to a small AI model, which rewrites them in strict notation, copying every number, never correcting it. The parser then checks that translation before anything is counted.

  2. 2

    Count

    Every stitch is defined by how many stitches it uses and how many it makes: k2tog uses 2 and makes 1, a yarn-over uses none and makes 1. Each row is walked stitch by stitch, for every size, and compared with the count the pattern promises.

  3. 3

    Fix

    When a row doesn't add up, Stitchproof tries small edits to that row's numbers, re-counts the whole pattern for each, and suggests the one that makes everything agree, or tells you the printed count is the typo.

We broke correct patterns on purpose.

Every number in six correct patterns (cast-ons, repeat counts, stitch counts and stated totals, in every size) was changed by ±1 and ±2, the typos designers actually make. Then Stitchproof proofed and fixed each broken copy on its own.

450
single-number typos, made in 6 correct patterns
100%
of the 423 typos that change a stitch count were caught
100%
of those were repaired so every row adds up again
89.4%
were restored to the designer's exact original text

The other 27 typos change no stitch count (for example, how many plain rows to knit), so no count-based check can see them. Recomputed on every build by lib/stitch/evaluate.ts; run it yourself with npx tsx scripts/eval-fixer.ts.

Who it's for

Designers
Check every size before you publish, and catch the errata your test knitters would have found on row 40. Send them a proof link instead of a spreadsheet.
Knitters
Check a pattern before you buy yarn, then knit it one row at a time with the count to expect and each row in plain English.
Buyers
Check a pattern before you pay for yarn. If the numbers can’t work, copy a plain-text report with the exact rows and send it to the seller or the marketplace.

Knit it once. Knit it right.

Stitchproof · built for Devpost's Build With AI: BasicsSample patterns are original and for demonstration.Source · MIT