From Prompt to Portfolio

A portfolio piece in 2026 is not the render. It is the documented trail of decisions that produced it.

Two designers answer the same brief. The first opens their portfolio to a single rendered logo with a one-line caption: brand identity for a local bike workshop, made with AI. The second opens to a brief, a moodboard where every reference is labelled with the quality it contributes, four prompt iterations each captioned with what changed and why, a critique pass with the strategic calls marked, the final piece, and a paragraph explaining what was chosen and what the model executed. A hiring manager reading both knows immediately which one has the judgment to brief: the trail gives them something to assess, and the single render gives them nothing to read but the surface.

That is the whole argument. A portfolio piece in 2026 is not a thing the designer made or a thing the AI generated. It is a documented decision-trail, and the documented judgment is the part a model cannot produce on its own. Recent hiring guidance lands in the same place: the strongest portfolios put the emphasis on decisions rather than polished final screens, and on how a designer uses AI as a creative partner without losing their judgment. The render is cheap now. The documented reasoning, for now, is not.

This article walks one project through the full Taste-First Method (build taste, generate, refine) and out the other side as a portfolio entry. Each stage is one of the other workflow articles, demonstrated rather than re-explained. The worked example is a brand identity for a fictional community bike-repair workshop. Everything shown is illustrative, manufactured to make the method legible, not a live client commission.

The brief, in one line: a volunteer-run bike-repair workshop wants an identity it can paint on a roller door, print on a worn enamel sign, and run on a one-page website. The audience is the neighbourhood, not an industry. The constraints are real: it has to read from across the street, survive being reproduced in one colour on a sticker, and feel like a working shop rather than a startup. The success criterion is recognition, not polish. The reader follows this one project from brief to documented entry. AI tools are used at every stage and disclosed in the final write-up, which is the point, not a footnote.

The brief is where the project earns the right to be specific later. Brief Writing with AI covers the framework; here it is filled in. The context dump names the workshop, its volunteers, the roller-door and enamel-sign applications, the one-colour reproduction floor, and the one non-negotiable: it must look like a working shop, not a tech brand. The audience is written down in plain terms. The brand voice is named in qualities, not adjectives: sturdy, hand-worn, unfussy.

A specific brief is what makes every later critique specific. Make it look good gives a model nothing. A one-colour mark that reads from across the street and survives on a sticker gives it a target. Most of the value in this stage is invisible in the final piece and load-bearing for everything after it. Skip it, and the model fills the gap with the average of everything it has seen, which is exactly what a hiring manager has already seen a thousand times.

Taste is the asset AI raises the price of, not the one it replaces. As Joshua Leigh argues, grounding the idea in Hume’s 1757 essay on taste, taste is the disciplined capacity for contextual judgement: knowing why a choice matters here, now, for this audience. A model can generate a thousand competent marks in a minute. Choosing the right one, and knowing why, is the work.

So the moodboard is not a wall of pretty pictures. Each reference carries a short label naming the quality it contributes: enamel-sign weight on one, single-colour confidence on another, hand-painted edge on a third. The references are gathered fast (a moodboarding tool such as Adobe Firefly Boards does the collecting), but the labels are the designer’s, and the labels are the taste. This is also where the stakes are clearest. Tommy Geoco names the risk phantom competency: the problem, as he describes it, is that AI augmentation makes output look more capable than the person behind it currently is. The named-quality moodboard is the first place the designer proves the competence is real, by showing they can say what is working and why.

This is the longest stage and the one the portfolio is really about. The asset is the workshop mark, and it takes four iterations, each captioned with what changed and why, in the vocabulary the Fundamentals pillar teaches.

Iteration one is deliberately vague (a logo for a bike-repair workshop) and comes back generic: a glossy chainring, a gradient, a typeface the model reaches for by default. Iteration two names the composition: a single chainring silhouette, centred weight, nothing else competing. Iteration three names the type: a sturdy grotesque (a plain, utilitarian class of sans-serif) set tight, not a rounded friendly sans, because the voice is hand-worn, not cheerful. Iteration four names the hierarchy and colour: the workshop name dominant, the strapline a third of its weight, the whole thing tested in one colour so it survives on a sticker. Same generator across all four (the prompt is the only variable). The render improves at each step, but the render is not what goes in the portfolio. The captions are. The trail shows a designer making decisions a model cannot make for itself, because the model does not know this workshop, this street, or this door.

The refine pass is the Zoom-In Method, covered in full in Design Critique with AI. The generated mark goes back to the model at 99% with a numbered self-review prompt: find the failures by category. The model returns surface notes (the strapline spacing is uneven, the one-colour version loses a detail, the mark is a few pixels off centre) and fixes most of them. In JM’s own practice running this loop, a model catches roughly seven in ten of its own surface mistakes this way, a working observation from doing the work, not a measured benchmark.

Then the 100% pass, which is the designer’s. The model handles whether the strapline is kerned evenly. The designer handles whether a chainring is too obvious a symbol for a shop that wants to feel hand-worn rather than clever, and decides to keep it because the neighbourhood will read it in half a second from a moving car. The model also suggested brightening the single colour for contrast; the designer overrules that, because a faded, slightly tired red is exactly the worn-shopfront feeling the brief asked for, and brightening it would make the place look new. Both of those calls get written down, including the one where the model was overruled, because a hiring manager learning to trust a designer wants to see the decisions that went against the easy option as much as the ones that went with it. The critique loop is where the surface gets cheap and the judgment stays expensive.

The documentation is the portfolio asset. The final entry, built here as an illustration of the method rather than a real commission, carries the brief, the named-quality moodboard, the four captioned iterations, the critique pass, the final mark shown on the roller door, a short reflection on what was kept and what was overruled, and a plain disclosure of which tools did what. As Ioana Teleanu puts it, AI will not make a portfolio meaningful on its own; the personal voice has to be put in by editing. The model can draft the case-study narrative. The designer edits it hard for voice, specificity, and decision clarity, and the editing is the authorship.

The reflection paragraph is where the entry earns its trust. It names the decision that did not work first time (the early chainring was too detailed to hold at sticker size, and had to be cut back to a silhouette) alongside the ones that did. A trail that only records the wins reads like a sales page; a trail that records a wrong turn and the correction reads like a designer.

Disclosing the AI use is part of the asset, not a confession. Transparency about the AI stack is an emerging norm, increasingly common rather than universal, and in the Australian context the honest disclosure habit is the one Client Disclosure describes. Tom Johnson, working in a different corner of the field, reaches the same instinct: he has the AI write a new folder full of markdown files with context, history, and constraints so the next build starts from the documented decisions. Different domains, the same instinct. The decisions are the thing worth keeping.

Hiring guidance through 2025 and 2026 is consistent about what gets read in a portfolio: real design thinking, with decisions justified by intent rather than asserted; clear ownership, stating what the designer owned and what changed because of it; authorship clarity, so it is obvious who made which call; and editing and voice, the evidence that a human shaped the work rather than forwarding the model’s output. A documented project answers all four in one artefact. A bare render answers none of them, because there is nothing to read but the surface.

Disclosure sits one notch softer. Being open about the AI stack is an emerging norm and good practice, not yet a universal requirement. The argument for it is not compliance. It is that the disclosure is where the authorship gets stated out loud, which is the thing being assessed anyway.

The failure mode this article exists to prevent is the AI-only loop: prompt, output, screenshot, portfolio. It is fast, it looks competent, and it is indistinguishable from everyone else who ran the same prompt. It is the phantom competency problem in its purest form, output without the understanding underneath it. The loop is genuinely quick: the Cynthia Liu case study, a non-coder building a portfolio site in four hours with AI, is the evidence for how quick. Speed is not the value. A portfolio of screenshots shows that a person can operate a model. A documented decision-trail shows that a person can design.

There is an older version of this distinction. The essay The machines are fine. I’m worried about us. tells it as a thought experiment from PhD supervision in astrophysics: two students do the same year of work, one struggling through it and internalising the judgment, the other delegating to an agent and shipping identical output with none of the understanding. The frame is a physics lab, but the pattern holds anywhere learning can be outsourced. From the outside the work is the same. The difference is what is left inside the person afterwards. Documenting your decisions is how you stay the first student. The screenshot loop is how you become the second.

Read the trail back and every captioned iteration is a fundamentals question answered out loud. The grotesque set tight rather than a rounded sans is typography. The single restrained colour that survives on a sticker is colour. The dominant name over the lighter strapline is hierarchy. The centred chainring with room to breathe is composition. The portfolio entry is the four fundamentals applied to a real problem and then written down. That writing-down is what turns a render anyone could prompt into a piece only this designer could account for.

The render takes a minute. The decision-trail takes the training. A Cert IV is where that training happens: the practice of making a decision, knowing why, and being able to defend it, which is the practice that survives the arrival of tools that render faster every month. The model will keep getting better at the output. It will not get better at knowing what this workshop, on this street, needed. That judgment is the portfolio, and the portfolio is the proof.

Next: build the trail one stage at a time. Brief Writing with AI, Moodboarding with AI, and Design Critique with AI are Stages 1, 2, and 4 in full. Build Taste, Generate, Refine is the method this article demonstrates end to end. Client Disclosure covers the disclosure habit.