Custom AI Assistants

Custom wizards allow you to save specific prompts for repeatable tasks (for example, generating question banks, rubrics, or activity sheets). The user sees a normal chat, but the base instructions are applied to each conversation.

What is a custom assistant (general idea)

A wizard is basically a fixed prompt + optional context that is reused across multiple conversations.

Most major platforms (ChatGPT, Gemini, Claude) offer ways to create wizards or "projects" with different names:

• GPTs (ChatGPT): shareable wizards with instructions, documents and optional capabilities.
• Gems (Gemini): saved instructions that you can activate in each chat. Especially useful for creating Socratic or guided tutors.
• Spaces (Perplexity): spaces with base instructions and files, which stand out because when shared with students they do not require them to register or log in, robustly complying with privacy regulations.
• Projects (Claude): shared context between conversations with base instructions and attached documents.

Frequent changes

Names, interfaces, and plan requirements change frequently. The important thing is the concept: saving a reusable prompt so you don't have to rewrite it every time.

How to create a wizard (general flow)

Although each platform has its interface, the basic flow is:

1. Defines base instructions (the prompt that will always be applied)
2. Optional: Add background documents (e.g., resume, rubric, glossary)
3. Optional: Activate extra capabilities (web search, image generation, code analysis)
4. Save and reuse

Common error when writing instructions

DO NOT write "Create a wizard that generates questions...".

Directly write the prompt that the AI ​​will see in each conversation:

❌ Incorrect: "Create a test question generator that..." ✅ Correct: "You are an expert in educational evaluation. When the user provides a topic, generate 10 multiple choice questions in GIFT format..."

Typical structure of a teaching assistant

Role and objective:
You are an instructional design assistant for middle school and high school/VET.
Your objective is [specific task].

Fixed context:
- Level: [specify if applicable]
- Constraints: do not use personal data; do not invent regulations.

Output format:
- [Fixed headers / table / JSON / GIFT]

Flow:
1. Ask clarifying questions (only the necessary ones).
2. Generate the proposal.
3. Include observable success criteria.

Key concepts (platform agnostic)

Concept	What does it mean in teaching
Base instructions	The prompt that will always be applied (your "template")
Context / Documents	Files that the attendee can view (e.g., curriculum, activity examples, rubrics)
Share	Being able to give the assistant access to other teachers at the center
Project / Session	Maintain context between multiple conversations on the same topic

Educational use cases

Scenarios where an assistant makes sense

• Question bank generator: fixed instructions for GIFT format + quality rubric
• Material adapter: converts a text to different levels (support/extension) while maintaining objectives
• Activity reviewer: validates that a proposal includes observable objectives, evaluation and attention to diversity
• Subject tutor: prompt with syllabus, style and restrictions (e.g., "use everyday analogies")
• Specialized translator: translates with area-specific terminology and consistent format

When NOT to use a wizard

• For one-off or exploratory tasks (use a normal chat).
• When you need to change instructions constantly (there is no point in "pinning" it).

Example: Moodle test assistant

chatGPT | GPT's

👤

Create test questions in GIFT format importable to Moodle, following these rules:

• They will always be multiple-choice questions with 4 answers and only one correct one.
• Incorrect answers will deduct half a point unless otherwise indicated.
• All answers must be of similar length.
• Rigor must be prioritized. The correct answer must be the only one that is correct. The others must be manifestly incorrect.
• Questions with incorrect answers because they are not "the best definition" of those given, "the main disadvantage" or other forms of graduation must be avoided. Incorrect answers must be manifestly incorrect to avoid ambiguities and questions must be clear.
• None of the answers should be trivially dismissible. Avoid the usual bias that the correct question is always one of 3 similar ones and there is another more absurd one.
• The expected difficulty level is for university master's degree students unless otherwise indicated.
• Do not include a number before the questions, they will be shuffled.
• Material will normally be provided to generate the questions, avoid making direct references such as "according to the notes". The questions must contain by themselves all the necessary information to answer them. Students will not be able to consult the notes during the test.
• Be careful avoiding biases such as the correct answer being the most detailed or the longest one.

An example of the desired question format would be the following:

---

// question: 91483439 name: What is stratified sampling? ::What is stratified sampling?::[html]What is stratified sampling?{ ~%-50%Dividing the dataset into groups and guaranteeing that all samples of a group are in the same set (test, validation, or training). =Dividing the dataset into groups and taking a percentage of each stratum maintaining the original proportion of the classes ~%-50%Dividing the dataset into groups and increasing the presence of minority classes by duplicating or generating new samples, which can lead to a more balanced representation of classes in the test set. ~%-50%Dividing the dataset into groups and increasing the complexity of the model, which can result in a more robust model but also harder to interpret and adjust. }

🤖

// question: 67895124 name: What is a generative model in artificial intelligence? ::What is a generative model in artificial intelligence?::[html]What is a generative model in artificial intelligence?{ ~%-50%An algorithm that classifies existing data into predefined categories based on features learned during training. ~%-50%A system that optimizes sequential decisions through interaction with an environment to maximize a cumulative reward. =A type of model that learns the distribution of training data and can generate new samples similar to these. ~%-50%An approach that reduces dimensionality of complex data by transforming them into simpler representations while preserving essential information. }

// question: 38215907 name: What is the main difference between supervised and unsupervised learning? ::What is the main difference between supervised and unsupervised learning?::[html]What is the main difference between supervised and unsupervised learning?{ ~%-50%Supervised learning uses neural networks while unsupervised employs traditional statistical algorithms. =Supervised learning uses labeled data as correct examples while unsupervised works with unlabeled data. ~%-50%Supervised learning requires less computational power while unsupervised needs more powerful computers. ~%-50%Supervised learning was developed in the 2010s while unsupervised has existed since the 1980s. }

// question: 29463187 name: What is cross-validation? ::What is cross-validation?::[html]What is cross-validation?{ ~%-50%A technique to reduce model complexity by removing redundant features from the dataset. ~%-50%A method to evaluate the performance of several models simultaneously on a single training set. ~%-50%A procedure that combines predictions of several models to improve the overall accuracy of the system. =A technique that divides the data into multiple subsets to train and validate the model, using each subset as test data once. }

This example shows how a well-designed prompt can become a reusable wizard for generating tests. More details in Tests for Moodle.

Checklist to create a good assistant

Before "saving" a wizard, check:

• Clear objective: what final product should you generate?
• Format defined: table / list / JSON / GIFT / Markdown with headers?
• Quality restrictions: success criteria? verifiability?
• Privacy: do you avoid asking for personal data from students?
• Clarification questions: Do you ask for what you need before responding?
• Reusable: does it work for multiple similar topics/situations?