Your Design Practice

What the last five lessons were really about

If you strip out the prompts, this course made one argument repeatedly.

A language model predicts plausible text. That makes it excellent at anything with a strong recognisable shape and high volume: enumerating states, drafting copy, structuring a flow, listing objections, writing the documentation nobody wants to write. It makes it dangerous at anything that depends on truth, because plausible and true are produced by the same process and arrive looking identical.

Everything practical followed from that. Trace every research finding to a real transcript line, because it will invent one. Treat the first flow as disposable, because it will hand you the median. Brief the voice with real strings, because its default voice is a performance of helpfulness. Withhold your preference before asking for criticism, because it would rather agree than be useful.

None of that is about a product or a feature. It is about a mechanism, which is why it should still be true after the tools change again.

The test for anything this course did not cover

You will meet a task next month that is not in these lessons. Ask three questions.

Is the work high in volume and low in irreducible judgement? Listing, restructuring, reformatting and drafting all qualify. Deciding what matters does not.

Can I check the output against something real? A transcript, a design system, a real user, a measurement, your own eyes on the screen. If there is nothing to check it against, you are not reviewing the output, you are being persuaded by it.

Would I be comfortable if a colleague saw exactly how this was produced? This is the ethical version of the same question, and it catches most of the cases the first two miss.

Two yeses and a no on the second question is the specific combination to be careful about. That is the shape of AI-generated research findings, AI usability feedback, and confident AI accessibility sign-offs.

What to never delegate

Some of these are practical limits and some are professional ones. Both matter.

Talking to users. Not the recruitment screener, not the discussion guide, not the summary. Those can all be drafted. The conversation itself cannot, and neither can the moment in a session where a participant does something surprising and you decide to abandon your script. A simulated participant produces what feedback sounds like, which is the most convincing wrong thing available to you.

The decision about what to build. A model asked to design a feature will design one. It has no way to conclude that the feature should not exist, that the team should fix the existing flow instead, or that this is the third time your organisation has tried this and the reason it failed twice was never technical.

Visual judgement. Hierarchy, restraint, rhythm, when to break the grid. There is no eye behind the output and no point of view about who is looking.

Anything you cannot verify. If the output is a claim about the world, and you have no way to check the claim, do not carry it forward. This applies with particular force to statistics, benchmarks and confident numbers. Those are exactly the outputs most likely to be fabricated and most likely to be repeated by someone else.

Material you are not permitted to share. Participant transcripts, recordings, anything covered by a consent form that did not mention third-party tools. The rule here is not about risk appetite. It is what you told the participant.

Responsibility. Your name goes on the work. "The AI produced it" is not available as a defence, and nobody has ever accepted it.

A realistic week

None of this needs to be a workflow overhaul. A version that actually holds up looks quite ordinary.

You still run discovery yourself. You paste your notes in with participant labels and ask for themes with verbatim quotes, then spend twenty minutes searching the transcripts for those quotes and delete anything that fails. You describe the flow you have in mind and ask for the states you forgot, then decide which four of the fourteen actually need designing. You write the important copy yourself and generate three options for the fifty strings you would otherwise have rushed. Before the review, you ask for the five strongest arguments against your decision and you prepare answers to the two that landed.

That is the whole thing. It is not a transformation, and any course promising one is selling you something. It is a few hours a week returned, and one more pass of scrutiny than you would otherwise have had time for.

Where to go next

Prompt Engineering if the briefs in this course felt like guesswork. It covers the general mechanics properly, and everything here gets easier afterwards.

Context Engineering if you noticed long sessions going strange. It explains why synthesis conversations start referring to their own earlier summaries instead of your transcripts, which is the single most likely way a checked finding becomes an unchecked one.

AI Safety, Privacy and Verification if the questions about participant data and consent are live at your organisation, or if you are the person other people ask about them.

AI for Graphic Design if you want the image generation side, which this course deliberately left alone.

Build a Website with Lovable if you have ever wanted to put a working prototype in front of someone rather than a clickable file.