When we use the Advanced mode with Tokens in the AI Automator settings, Tokens which are not resolved by the Token system stay in the AI Prompt. This is not much text for AI, but everything what can be avoided to send to AI, should be removed. As an example. I have an Entity where I defined a textarea called field_user_prompt which can optionally (!) be used to alter the prompt I setup in the field AI automation. So there is a Token like this: [node:field_user_prompt:value]. This stays when the user left the field empty. It's only replaced when something is filled in. I do think that unresolved tokens have no benefit to a prompt which gets send to AI, so why not have a clean up anyways. ` public function renderTokenPrompt($prompt, ContentEntityInterface $entity) { // Get variables. return $this->token->replace($prompt, [ $this->getEntityTokenType($entity->getEntityTypeId()) => $entity, 'user' => $this->currentUser, ]); } ` should be ` public function renderTokenPrompt($prompt, ContentEntityInterface $entity) { // Get variables. return $this->token->replace($prompt, [ $this->getEntityTokenType($entity->getEntityTypeId()) => $entity, 'user' => $this->currentUser, ], ['clear' => TRUE]); } ` <details> <summary>Contextual information for getting started</summary> ## Context AI Automators (`ai_automators`, a submodule of the AI module) let a site builder configure a field on an entity to be filled automatically by an LLM when the entity is saved. The part this change touches is the **Advanced prompt** you set per field in the field's AI Automator settings, which is a template that can embed tokens like `[node:field_user_prompt:value]` to splice in other field values. When a token can't be resolved — e.g. an optional field left empty — `Token::replace()` leaves the raw token text in the prompt. The fix adds `['clear' => TRUE]` to `Token::replace()` in `renderTokenPrompt()` so unresolved tokens are stripped instead of being sent to the model. ## Why does this matter Every unresolved token is shipped to the model verbatim: it wastes input tokens (and money) on every run and adds noise that can confuse the model or leak internal field machine-names into the prompt. Site builders reasonably expect an empty optional token to disappear, not to reach the LLM as literal `[node:…]` text. ## How to test Assumes an AI provider + Chat model are configured and you're on your MR branch. 1. Enable `ai_automators` plus an observation aid — either `ai_logging` (read the exact prompt sent) or `ai_test`'s Echo provider (returns the prompt as the field value, easiest to read). 2. On a content type, add an optional `field_user_prompt` (text) and a target `field_ai_target`; configure an AI Automator on the target with an Advanced prompt embedding the token, e.g. `Extra user instructions: [node:field_user_prompt:value]`. Run it directly on save. 3. Save a node with `field_user_prompt` **empty**: with the fix inactive the prompt shows the literal `[node:field_user_prompt:value]`; with the fix active (and the opt-in checkbox on, if you built one) the token is gone. Then fill the field in and confirm it still substitutes normally. **Further reading:** [AI Automators](https://git.drupalcode.org/project/ai/-/blob/1.x/docs/modules/ai_automators/index.md) and [AI Prompt Management](https://git.drupalcode.org/project/ai/-/blob/1.x/docs/developers/ai_prompt_management.md) in the project docs. ## Acceptance With the change applied, unresolved/empty-optional tokens no longer appear in the prompt sent to the model, while tokens that do resolve still substitute correctly (no regression). Note for the MR: `clear => TRUE` strips *all* unresolved tokens in the prompt, not just the empty field — flag whether that should be the default or sit behind the "(Checkbox?)" opt-in. </details>