Learn AI by doing real design work.

Start with the job you need done. Use a general assistant for thinking and writing, a research tool when sources matter, an image tool for visual exploration, and a coding agent when you need to change working software. You do not need every tool. Choose one primary tool, learn its strengths, and add another only when the task calls for it.

A model is the underlying system. A prompt is your instruction. Context is the material available to the model during the task. Multimodal means it can work with more than text, such as images or audio. A hallucination is a confident but unsupported output. An agent can take a goal, use tools, and complete several connected actions on your behalf.

Write prompts in five parts: state the outcome, provide the relevant context, name the constraints, describe the artifact you want, and define how it should be evaluated. If the result is weak, improve the missing part instead of adding vague instructions like make it better. Save prompts that produce a useful repeatable workflow.

Use AI to increase the number of directions you can explore and the speed at which you can test them. Do not outsource the decision itself. Ask what evidence supports a recommendation, compare alternatives, inspect the actual output, and be ready to explain why the final choice serves the user and the product.

Use a screenshot, design file, or live page with enough detail to understand the system rather than a single isolated component. Ask AI to inventory the grid, type scale, spacing, color, imagery, responsive behavior, and interaction patterns before it writes code. Confirm what should be reproduced, what should be adapted, and which assets are still missing.

A useful build prompt combines the reference analysis with real content and a clear scope. Tell the agent which project it is working in, what it may change, which breakpoints and states matter, and how you will judge fidelity. Let it produce a working first pass, then review the result in the browser instead of debating the code line by line.

Ask the agent to identify the project framework and package manager, install only what is required, and start the existing development command. Open the local URL it reports, usually an address such as localhost with a port number. Leave that terminal process running so saved code changes refresh in the browser, and stop it when the session is finished.

Choose a slice with a clear boundary and enough importance to prove the direction. Describe what should feel different, which elements must remain, and the evidence you will use to approve it. Ask the agent to limit its edits to that slice, review the live result at relevant sizes, and iterate until the pattern is genuinely ready to repeat.

Before scaling, name the rules that made the slice successful: typography, spacing, component structure, behavior, or content pattern. Ask the agent to find every true instance, list likely exceptions, and apply the change systematically. Review representative pages plus the exceptions, then check the change set for unrelated edits.

Test the full path a visitor will take, not just the section you edited. Check small, medium, and large viewports; keyboard focus and interactive states; readable contrast and type; real content lengths; links and media; loading and error behavior; and consistency with the source design. Let AI find defects, but personally verify every fix in the rendered experience.

Select examples that answer a real question rather than simply looking impressive. Include work aimed at a similar role or audience, one example that sets a higher bar, and one that solves the communication problem differently. Save the exact pages you want evaluated so the comparison stays focused and reproducible.

Write the criteria before reviewing examples so the rubric does not bend toward whichever portfolio you like most. Define what evidence would indicate strong or weak performance for each criterion. Prefer notes tied to visible page content over unexplained numerical scores, then use the rubric consistently across every example and your own site.

Give each simulated reviewer a distinct goal and limited attention. A recruiter may scan for fit and clarity, a design leader may probe craft and decision quality, and a product partner may look for collaboration and commercial thinking. Compare where their feedback overlaps and treat persona-specific conclusions as hypotheses to validate with real people.

Ask AI to place each conclusion into one of three groups: direct observation, reasonable inference, or unanswered question. Require the exact page, passage, or visible element behind an observation. Open the source yourself before acting on anything important, and discard claims that cannot be traced back to evidence.

Create a reusable context document containing the roles you want, your current resume, the job description, your portfolio, interview stage, constraints, and what worries you most. Tell the coach how direct to be and which facts it must never invent. Update this context as you learn more so each practice session builds on the last.

Collect stories before trying to perfect the wording. For each one, record the situation, your responsibility, the decision or action you personally took, the result, and what you learned. Tag stories by competencies such as craft, influence, ambiguity, conflict, leadership, or failure so you can retrieve the right example quickly under pressure.

Begin with a realistic sequence of questions delivered one at a time. Answer aloud and have the coach withhold feedback until the mock ends so the rhythm resembles a real interview. In later rounds, add follow-up questions, skepticism, shorter time limits, or a question that does not match your prepared stories.

After a mock or real interview, first write down how you think it went. Then compare that impression with the transcript and evaluate substance, structure, relevance, credibility, and differentiation. Choose one recurring bottleneck, rewrite or rehearse the weakest moment, and carry one specific behavior into the next session.

AI can make every answer sound smooth and interchangeable. Keep the details only you would know, the tradeoffs you genuinely considered, and the opinions your experience has earned. Read every suggested answer aloud, remove language you would not use naturally, and practice from ideas rather than memorizing a script.

Store the inputs and process, not only the polished result. Keep the prompt, reference material, useful context, intermediate decisions, rubric, and final artifact in one clearly named folder or document. Add a short note explaining when to use the workflow again and what you would change next time.

Before sending material to a model, ask whether you have permission to share every part of it and whether the chosen tool may retain it. Replace sensitive details with realistic placeholders when possible. Before publishing, verify sources, remove private information, confirm usage rights for assets, and take responsibility for every claim in the final work.

Compare the workflow with how you would have approached the task without AI. Note where it saved time, expanded the option space, exposed a blind spot, or introduced extra verification work. Identify the moments where your taste, product knowledge, or ethical judgment changed the result, since those are the capabilities worth strengthening.

Choose a task small enough to finish but real enough to matter. Define the artifact you will produce, the AI role, the evidence needed for approval, and the time you will spend. Schedule a short retrospective afterward and save what worked so the experiment becomes the start of a repeatable practice.