Prompt practice
Turn weak prompts into structured instructions with context, constraints, output formats, examples, and evaluation rules.
Agentic sandboxVerified practice layer
Practise AI skills before they reach production workflows.
The DeepMind Resources sandbox turns lessons, AI updates, and tool intelligence into guided practice. Work through prompts, workflow simulations, RAG exercises, tool-calling tasks, agent design, and validation drills inside a structured learning environment.
Sandbox Guide AgentPublic preview
Meet the first guided sandbox agent.
This guide helps users practise with safe example scenarios before entering the full workspace. Signed-in users will later use guided tasks with saved attempts, feedback, progress records, and business-ready practice routes.
Use example or anonymised scenarios only. Keep customer details, employee information, passwords, contracts, financial data, and confidential company material out of this public demo.
Prompt Coach
Improve a weak AI prompt
Rewrite the prompt so an AI system knows the audience, purpose, constraints, output format, and quality standard.
Demo starter
Write me something about AI for small businesses.
Agent readout
Waiting for input
Choose a route, add a short example, then open the guide to see how this sandbox task should be improved.
What the sandbox covers
A structured practice layer for modern AI work.
DeepMind Resources treats the sandbox as a skill system, not a toy demo. Each task is built around a scenario, a practical goal, success criteria, and a review checklist so users can practise deliberately.
Choose a route
Start with prompt design, workflow mapping, tool calling, RAG, agent structure, or validation practice.
Work from a scenario
Every exercise begins with a clear situation, safe sample input, and a practical goal instead of open-ended play.
Review the output
The sandbox helps users check structure, evidence, risk, privacy, assumptions, and whether the result is ready for real work.
Workflow simulations
Map real tasks into repeatable AI workflows with clear inputs, review points, outputs, and human approval steps.
Tool-calling exercises
Practise when an AI system should call a tool, what data it should pass, and how to validate the result before action.
RAG and knowledge tasks
Work through retrieval, source grounding, answer quality checks, and knowledge-base design using safe sample scenarios.
Agent design tasks
Design bounded agent workflows with goals, tools, permissions, stop conditions, logs, and review gates.
Validation drills
Identify weak claims, hallucinations, missing sources, privacy risks, and unsafe assumptions before outputs are trusted.
Prompt improvement task
Rewrite a vague request into a clear prompt with role, task, context, constraints, and a measurable output format.
Tool selection scenario
Compare two AI tools for a defined business or learning task, then explain the trade-offs and safe next step.
Agent workflow design
Design a controlled agent workflow with tool boundaries, approval checkpoints, failure handling, and evaluation criteria.
Workflow readiness exercise
Convert a workplace task into an AI-assisted process with privacy rules, staff responsibilities, and review standards.
Skill evidence
The goal is practical competency, not page views.
Sandbox tasks are designed to help users prove they can apply a skill. A completed task should show more than passive reading: it should demonstrate judgement, structure, validation, and safe execution.
Every task practises a real AI skill.
Every task has a clear output and success criteria.
Sample data is used before sensitive business data.
Validation, privacy, and human review are part of the workflow.
Sandbox practice connects back to lessons, tool intelligence, and progress tracking.
Access path
Start with the learning path, then move into guided practice.
The public sandbox page shows how guided practice works. Full sandbox tasks belong inside the signed-in DeepMind Resources workspace so attempts, progress, feedback, and business training context stay private to the right account.
Protected practice, public explanation.
Explore the public preview here, then use the signed-in workspace for private attempts, saved progress, team tasks, and guided sandbox feedback.