Skip to main content

AI Training and the Playground

AI Training is where you tell the AI how your store speaks and what it must never get wrong. Everything here is store-level context: it travels with every enrichment of that store — products, categories, collections, pages and FAQs — from the next run after you save.

Open a store and click AI Training in the store menu. Only owners and admins see it.

The page has up to five tabs:

TabWhat it holds
GeneralYour voice sheet (store instructions) and the brand documents it is built from. See The voice sheet below.
DefaultsFixed facts the AI should never guess: brand name, market, title format, keyword and category blacklists, length preferences. See Store defaults.
LLM ConfigurationWhich AI provider and model enrich your stores. You can add your own API key (OpenAI, Anthropic, Gemini or Perplexity); without one, the platform default is used. This setting belongs to your organization, so it applies to all its stores.
PromptsThe prompt templates per entity type, with a tester. Each Run test is a real AI call and uses one credit.
PlaygroundA bench to try changes before they touch your catalog. Not open on every store yet — see The Playground.

Nothing on this page writes to your storefront. It changes how the next enrichment is written; publishing is still a separate step you take.

The voice sheet​

The Store instructions card on the General tab holds your voice sheet: the brand rules the AI reads before every enrichment. Voice, tone, terminology, claims to avoid, facts to keep consistent.

You can type it, or build it from your own brand material.

Upload brand documents​

Click Upload file and choose a brand manual, a voice guide or a glossary.

  • One file at a time, up to 10 MB.
  • Formats: .txt, .md, .pdf and .docx.
  • Clione keeps the document's text and discards formatting, images and tables. A scanned PDF with no text layer has nothing to read and is refused.

Uploading a document doesn't change your sheet. The documents stay on Clione's side and are never sent to the AI as they are: they feed the sheet, and the sheet is what travels with every enrichment. Documents don't count towards the sheet's size limit. You can remove a document at any time.

Generate the sheet from your documents​

Click Generate sheet from documents. The button shows what it will cost before you click: one credit for each started 50,000 characters of document text, with a minimum of one. If the generation fails, nothing is charged.

The AI reads your documents and writes a sheet with the sections they support — voice and tone, audience, terminology to use and to avoid, claims to avoid, facts to keep consistent. It is told to use only what your documents say, and to keep the rules already in your sheet unless the documents contradict them.

Nothing is replaced until you say so. You get a proposal side by side with your current sheet, with the lines added and removed marked. Choose Replace sheet or Keep my sheet. If you edited the sheet by hand since it was last generated, the dialog warns you that replacing it drops those edits.

Once it's applied, the card says when the sheet was generated. If you later edit it by hand, or your documents change, the card says so too — regenerate when you want the sheet to catch up.

Edit it yourself​

The sheet is plain text. Edit it and click Save instructions; saving by hand is free. The Memory available ring shows how much of the sheet's limit is left (10,000 characters). You can't save while you're over it, because every character of the sheet is sent with every enrichment.

Suggested lines​

If you write a rule into one entity's prompt that reads like brand voice, it shows up under Suggested for your sheet. Accept adds it to the sheet, Reject discards it. Nothing is added without your click.

Use my documents per entity​

Optional, and off by default. With Use my documents per entity on, the few excerpts of your documents most related to each entity travel with its enrichment, as well as the sheet. If an excerpt and the sheet disagree, the sheet wins. The switch needs at least one uploaded document, and it isn't available on every installation; when it isn't, the card says so.

The Playground​

The Playground is a calibration bench. Pick one entity, change one thing — the voice, the model, the prompt, or a one-off instruction for this run — and run the same enrichment pipeline your store uses. The results stack side by side so you can compare versions before you change anything for real.

It writes nothing​

A Playground run is a dry run:

  • Nothing is written to the entity. No new version, no change to its meta title, description or FAQs.
  • Nothing is written to your storefront.

The only way something leaves the Playground is Save to store, which makes the winning voice, model or prompt your setting: the voice and the prompt for this store, the model for your organization. It writes only what you changed in that run and needs an owner or admin. It still writes nothing to the entity and nothing to your storefront: the change applies to the next enrichment you start.

Every run spends a credit​

Each run is a real AI call. Every run uses one enrichment credit from your organization's monthly allowance, whether it's an enrichment or an FAQ generation. A run that fails is not charged. When your credits run out, runs are refused until next month, an upgrade or an Expand credits add-on.

On top of that, each store has a daily limit on Playground runs. The counter next to the run button shows how many you've used today; the limit resets at midnight UTC.

How a run works​

  1. What are you testing? Pick an entity type — product, category (BigCommerce) or collection (Shopify), or content page — and search for the entity.
  2. Choose the task: enrichment, or FAQ generation.
  3. Change what you want to try. You can edit Voice & tone (your sheet, for this run only), the LLM config (provider, model, temperature), the editable part of the Prompt, and Steer this run (for example, "lead with the 2-year warranty"). The answers from Teach the AI and the rest of the context the pipeline adds are shown read-only.
  4. Click Run enrichment or Generate FAQs.

Each run shows the meta title, meta description, keywords, enriched description, quality scores and reasoning (or the generated FAQs, with how many passed Clione's checks), plus a Run recipe listing exactly what it ran with. The Playground can also open full screen in a new tab.

Who can use it​

The Playground is being opened store by store; access is granted per store and per person by Clione. If it isn't open for you yet, you'll either not see the tab or see it greyed out as Coming soon. Ask through Report a problem in the top bar if you think you should have it — see Support cases.