shopify sidekick guardrails localization sku catalog data-accuracy

How to Keep Shopify Sidekick from Changing Your Store's Language or SKUs

Sidekick can quietly drift back to English on a multilingual store, or invent a SKU it doesn't actually have. Here's how to lock both down.

4 min read SidekickSkills View as Markdown

Shopify Sidekick can silently revert multilingual content to English mid-batch, and it can invent SKU or spec values that don’t exist in your store — both without telling you. Neither is a bug you wait out; both are fixed by adding an explicit instruction before you run the task, not by hoping the next prompt behaves.

Here’s what’s actually happening and how to lock it down.

The language problem: Sidekick drifts back to English

If your store’s content language is set to Spanish, French, German, or anything other than English, you’ve probably seen this: Sidekick starts a batch of product descriptions or meta fields correctly in the right language, then a few items in, one slips back to English. On a long batch it can happen more than once, and it’s easy to miss if you’re not reading every item.

This isn’t Sidekick asking you which language to use — it’s assuming, and English is its default assumption unless you override it.

What actually fixes it:

  • State the target language explicitly at the start of the prompt, not just once at the top of your store settings.
  • Ask Sidekick to confirm the language back to you before it starts generating.
  • For multi-market batches, have it label which language each output is in so a mismatch is visible at a glance, not buried in paragraph three of item 12.
  • Ask it to re-check its own output before showing you the draft — a self-check catches a lot of what a first pass misses.

If you skip this, the failure mode is quiet: you get a batch that looks done, publish it, and a customer (or you, weeks later) finds the one product still in English.

The SKU problem: Sidekick can invent values that don’t exist

Ask Sidekick to “fill in missing SKUs” or “generate barcodes for products that don’t have one,” and it can produce values that look right — plausible formatting, consistent-looking codes — without those values being grounded in anything in your actual store data. The same risk shows up with specs: a wattage, a Bluetooth version, a dimension stated with total confidence that isn’t sourced from the product’s real fields.

This matters more than a wrong description, because a fabricated SKU or barcode can break inventory matching, fulfillment, or a sync with an app that depends on that field being accurate.

What actually fixes it:

  • Require Sidekick to pull existing SKU/barcode/spec data from your store records first, and never skip an active product just because its current data looks messy.
  • Tell it not to overwrite or “correct” an existing value on its own — flag it and ask you first.
  • Only generate new values for fields that are genuinely empty, using a pattern you’ve confirmed (e.g. [YOUR PATTERN, e.g. "CATEGORY-COLOR-SIZE"]), not a scheme it invents.
  • Ask for a table before anything is written: product, field, current value, proposed value, and whether that value came from real store data or a pattern you approved. Anything Sidekick isn’t confident about should be flagged, not guessed.

The underlying pattern: make it show its sourcing

Both problems above are versions of the same thing — Sidekick presenting output with more confidence than the underlying data supports. The general fix applies beyond language and SKUs: whenever you’re asking Sidekick a question you plan to act on (a diagnostic, a recommendation, a “why did X happen”), have it label each claim as VERIFIED — pulled directly from your store’s live data, source named — or INFERRED — a pattern, a guess, or general best-practice knowledge.

This also helps when you push back on a wrong answer. Without this rule, Sidekick can flip straight from “yes, you can do this” to “actually, Shopify doesn’t allow it” with no new evidence in between. With the rule in place, you’re asking it to say plainly what it’s uncertain about and point you to where you can verify it yourself, instead of just reversing its story.

Putting it together

None of this requires a different tool — it requires being explicit up front instead of correcting after the fact:

  1. Multilingual content: name the target language, ask for confirmation, and require a self-check pass before you see the draft.
  2. SKUs, barcodes, specs: require sourcing from real store data, no invented numbering schemes, and a before-you-apply table showing where every value came from.
  3. Anything you’ll act on: ask for VERIFIED vs. INFERRED labeling so you’re not treating a guess as a fact.

Add the relevant instruction to the top of your prompt each time the task touches one of these areas, and review the batch before you apply it — Sidekick is a strong first pass, not a final check.

Frequently asked questions

Why does Shopify Sidekick keep reverting my content back to English? ▾

Sidekick doesn’t reliably hold a non-English language for an entire long batch on its own. It’s common for it to start correctly in your store’s set language and then drift back to English partway through a large request, especially on multi-item batches. You have to explicitly instruct it to check the target language before generating and re-check its own output before showing it to you.

Can Sidekick invent a SKU if one doesn't exist yet? ▾

Yes, if you let it. Asked to “fill in missing SKUs,” Sidekick can generate plausible-looking values that aren’t grounded in your actual store data or numbering scheme. The fix is to require it to pull real variant data first, only generate for genuinely empty fields, and show you where every value came from before you apply anything.

Does Sidekick tell me when it's guessing versus using real store data? ▾

Not unless you ask it to. By default, Sidekick can present an inferred or generic answer with the same confident tone as a fact pulled from your live data. Add a standing rule that it label every claim VERIFIED (with the exact source) or INFERRED, so you can tell the difference before you act on it.

Do these guardrails slow down every Sidekick request? ▾

No — add them only to tasks where the risk is real: multilingual content batches, SKU/barcode/spec generation, and any diagnostic or “why is X happening” question you plan to act on. For a single English product description or a quick theme tweak, you don’t need the extra instructions.