Choose one job, such as turning meeting notes into action items. Say who the result is for, what it must include and how it should be presented. Before asking, write down what would make the answer correct.

Try it: Use a short set of notes you can judge yourself. Compare “summarize this” with a request for actions, owners and supporting sentences. Ask the AI to mark missing information rather than guess.

Does following your requested format make an answer correct?

No. An answer can have the right headings and still invent an owner or miss a decision. Check the contents against your source and the success criteria you set.

Try the prompt exercise →Write a useful prompt →

2. Give the AI the material the task needs

Instructions explain the job; reference material provides information for that job. Label documents and versions, include the relevant passages and ask for evidence. A longer conversation or a larger document does not make an unclear request precise.

Try it: Choose a document question whose answer you already know. Compare the full document with just the relevant section. If an answer misses the evidence, supply the exact paragraph and try the same question again.

If an answer misses a document detail, should you rewrite the whole prompt?

First check whether the relevant text reached the AI. Supplying the exact passage is a useful control experiment: it helps you distinguish a source-selection problem from trouble using the evidence.

Try the context exercise →Choose useful context →

3. Verify the parts that matter

Separate factual claims, calculations and suggestions. Check claims against original sources, recompute important totals and compare extracted details with the document. For AI news, distinguish an announcement, a paper and a result you can actually use.

Try it: Highlight the factual claims in one AI-generated paragraph. Open its sources and label each claim supported, contradicted or unresolved. Keep uncertainty visible when you revise the paragraph.

Is asking the same AI “are you sure?” an independent check?

No. It may produce another explanation without new evidence. Use the original source, a separate calculation or an appropriate test. Agreement between answers alone does not establish correctness.

Try the verification exercise →Check an AI answer →

4. Turn a useful result into a repeatable process

Keep a few representative tasks, their source material and the expected result. Record the prompt and how much correction the answer needs. Change one part at a time, then rerun those examples when your prompt, model or app changes.

Try it: Save an easy example, an awkward example and one with missing information. Compare your original prompt with one revision on all three. Keep the version that needs fewer corrections for your work.

Does one impressive answer show that your workflow is reliable?

No. Try several examples, including awkward and missing-information cases. Compare correctness and the work needed to fix the output, then revisit the checks when the task or tools change.

Practice changing one thing →Build a record of useful checks →

5. Explore running a model on your own hardware

Running AI locally adds a setup decision to the process of choosing a task and checking its result. A model file contains the learned weights; a runtime loads and runs it. Your hardware needs room for the weights, conversation and software, and you still need to judge the answers.

Try it: Open the local AI calculator, choose your hardware and start with a modest context. Inspect where the memory goes. Follow the local learning path to explore precision, context and a first run.

Does fitting a model into memory tell you how useful its answers will be?

No. Memory fit, runtime compatibility, speed and answer quality are separate checks. Reuse the representative tasks and evidence checks you learned here when trying a local model.

Start the local AI path →Explore the local AI tools →

Terms you will encounter

open-weight language models: Models whose learned weights are available to download. Their licenses still set the terms for use and redistribution. See the lesson →
tokens: A piece of text a model reads or generates. It may be a word, part of a word, punctuation or whitespace. Different models can split the same text differently. See the lesson →
context window: The working material available during a response: instructions, conversation history, supplied documents and the answer being generated. See the lesson →
model weights: The learned numerical values loaded from a model file. These are different from the temporary information saved for a conversation. See the lesson →
stored parameters: The learned values in a model. A label like 8B describes roughly eight billion values, not eight gigabytes. See the lesson →
weight precision: The numerical detail used to store model values. Quantization uses fewer bits per value: Q4 is roughly four bits, Q8 eight, and FP16 sixteen. See the lesson →
conversation memory: Saved attention information and, for some models, recurrent state reused while generating the next token. See the lesson →
VRAM: Working memory on a dedicated graphics card. It is separate from system RAM and from disk storage. See the lesson →
unified memory: A memory pool shared by the CPU, GPU, operating system and other applications. See the lesson →
runtime overhead: A runtime is the software that loads and runs a model. Inference means using the trained model to produce output. See the lesson →
GiB: A gibibyte is 1,073,741,824 bytes. A decimal gigabyte (GB) is 1,000,000,000 bytes. See the lesson →
system reserve: Available memory is the estimated capacity left after reserving space for the system. Headroom is what remains after the workload is loaded. See the lesson →
prompts: The instructions or input you send to a model to ask it to do something. See the lesson →
GGUF: A model file packages the learned values and metadata for software to load. GGUF is a common file format; MLX is a framework with its own model ecosystem. See the lesson →
bandwidth: How much data can be moved per second, such as from memory to a processor. See the lesson →
concurrent requests: The number of conversations being processed at the same time. See the lesson →