Who is deciding where this conversation goes?
Mixed-initiative interaction
A person and a system can each propose actions or change the direction of a task. Good design makes those proposals visible and lets the person intervene.
LevrAge Atlas / A field guide to working with AI
Name what’s happening.
Choose what to do next.
When an assistant loses the thread, invents a detail or steers the plan, there are useful ideas behind the experience. Start in plain language. Learn the formal name. Try it on a case.
Free to explore. No account, payment or research participation required.
Start with a question you recognize.
Park means save a concept for later on this browser. Only concept IDs are stored; no account or research signup. Other people using this browser can see the list. Clearing browser storage removes it.
16 concepts shown
Did the wording quietly change my decision?
Framing effects
Different descriptions of the same choice can change what people prefer. Which outcomes are foregrounded matters.
Why does an early assumption keep shaping the work?
Path dependence
Earlier choices can alter the options, incentives or costs of later choices. History can matter even when better alternatives remain.
Why does the same explanation land differently for different people?
Recipient design
People shape what they say for a particular listener, using assumptions about that listener’s knowledge and situation.
Did it answer the words or help with the purpose?
Gricean relevance
Grice’s account of cooperative conversation includes the expectation that a contribution is relevant to the exchange. Context and implication matter alongside literal words.
Why can an answer sound right and still be made up?
Confabulation / hallucination
Generative AI can produce false or unsupported content with the appearance of a factual answer. Fluent language does not establish truth.
Where did this claim come from?
Provenance
Provenance records the people, activities and source material involved in producing something. It helps you inspect how a result came to exist.
If it searched my files, why can it still miss things?
Retrieval-augmented generation (RAG)
A system retrieves material and gives it to a model to help produce an answer. The retrieval and generation stages can each fail.
Why did it lose track of something I already told it?
Context window
A model works with a bounded amount of input at a time. A product may select, retrieve or summarize previous material before it reaches the model.
Can a document give my assistant unwanted instructions?
Prompt injection
Untrusted content can contain instructions that try to redirect a model away from the authorized task. In indirect attacks, those instructions arrive through material the system reads.
Am I accepting it because it is useful, or because it sounds authoritative?
Automation bias and appropriate reliance
People can rely too heavily on automated advice, or fail to use reliable help. The useful target is reliance that matches the system’s actual capabilities in the situation.
Does “90% confident” mean it is right nine times out of ten?
Calibration
Probabilistic confidence is calibrated when predictions given a particular probability are correct at roughly that frequency across comparable cases.
How would I know this tool works for my actual task?
Evaluation
Evaluation compares a system’s behavior with explicit criteria on a defined set of tasks. A single impressive example gives little information about failure rates.
How much access does this assistant actually need?
Least privilege
Give a person or program only the permissions needed for the task. Reduce the damage that a mistake or compromise can cause.
Is the assistant checking my idea, or just agreeing with me?
Sycophancy
An assistant may adapt its answer to match a user’s stated beliefs rather than the evidence. Agreement can feel helpful while weakening a decision.
Is it following a recipe, or choosing its next action?
Agents and workflows
A workflow follows a designed sequence. An agentic system gives a model some control over which steps and tools to use. Products use these labels differently.
Try it on a fictional case
A useful word should change what you do.
Four short practices. Your choices stay in this page’s memory and disappear when you leave. They are not sent to the studio, used for research, or scored as a qualification.
Whose idea was this?
You: “Help our team reduce missed appointments.” Assistant: “Let’s build a subscription app.” You: “What should the app cost?”
When the answer is sound
Fictional order: 4 boxes at $20 each. Delivery is included. Summary: “The total is $80 including delivery.”
A detail the source never gave
Fictional order: 4 boxes at $20 each. Delivery is included. Summary: “The total is $80, and the boxes arrive tomorrow.”
Did AI actually save time?
A fictional task takes 15 minutes manually. With AI: 3 minutes to prepare, 2 to generate, 8 to check, and 6 to repair.