05 · Anti-slop: critique and the linter

"You can tell it's AI-made" has specific ingredients. OpenDesign fights it at three gates: before generation, after generation, and before the file lands.

The three gates

Gate	Where	What it does
Discovery form	Before	Forces you to state audience, tone, brand, and whether the data is real
Self-critique	After	Scores five dimensions, produces an evidence-backed report
lint-artifact	Before the file lands	Hard checks — a hit is a regression

Each covers a different stretch. Skip the first and the other two can only patch something built on a weak brief.

Gate one: the discovery form

The questionnaire before generation. It solves the "container soup" problem — with no context, a model can only give you generic containers: a hero, three cards, a CTA.

Get started chapter 5 covers filling it in. One point worth repeating: a long form is a signal, not an annoyance. It's telling you how much of the brief is still unstated.

Gate two: self-critique

After generation, artifacts get scored before you see them. Five dimensions:

Dimension	The question
Philosophy	Does this design have a position, or is it a stack of trends?
Visual hierarchy	Does the eye land in the right order?
Detail	Are edges, spacing, states, and micro-interactions actually handled?
Functionality	Does it work, or does it only look like it works?
Innovation	Is there judgement beyond the template?

Each scored 0–10, producing a self-contained HTML report with a radar chart, evidence-backed scores, and three lists: Keep, Fix, Quick wins.

The UI usually has a Critique entry point to run it manually. Running it after a first version beats staring at the artifact — it names what you've already stopped noticing.

Or ask for it in prose:

Run a five-dimension review on the current artifact: philosophy, visual
hierarchy, detail, functionality, innovation. Score each 0–10.

Every deduction needs concrete evidence — which element, which spot.
No generalities.

Output three lists:
- Keep: what's right, and why
- Fix: what must change, ordered by severity
- Quick wins: anything fixable in five minutes

Don't make the changes in this pass. Show me the report first.

That last line leaves you a decision point. Some deductions are its aesthetic preference, and you needn't agree.

Gate three: lint-artifact

This one is hard. The daemon's lint-artifact checks a subset of the anti-slop rules, and a hit is a regression, not a style preference.

The craft file is honest about the boundary: what's auto-checked, and what's advice, with the advisory parts marked "(guidance, not auto-checked)". That candour is more useful than a claim of total coverage — you know what to rely on and what to inspect yourself.

The actual ingredients of slop

What the project explicitly names, plus what shows up in practice:

Ingredient	How it appears	What to say
Purple-blue gradient	A gradient wash on the hero	"No gradient backgrounds. Use a solid from the system."
Generic icons	A row of line icons that mean nothing	"Icons must map to a specific function; prefer words."
Invented metrics	"10,000+ users", "99.9% satisfaction"	"Without real data, use a labelled placeholder."
Three cards	Every kind of content becomes three columns	"Let the content decide the layout. Three columns is not the default."
Uniform density	Every screen carries the same amount	"Allocate whitespace by importance."
Decorative imagery	Images unrelated to the argument	"An image explains a point or it goes."
Empty copy	"Empower", "end-to-end", "seamless"	"Copy states what it concretely does."

A self-check you can paste

Audit the current artifact against this list. Give a verdict and
evidence for each:

1. Any gradient backgrounds? If so, is the gradient defined in the
   design system?
2. Does every icon map to a specific function, or are some decorative?
3. Are all numbers real or illustrative — and are the illustrative
   ones labelled?
4. Is the layout content-driven, or defaulted to three columns?
5. Is density allocated by importance, or uniform throughout?
6. Does each image explain something?
7. Any content-free words in the copy ("empower", "end-to-end",
   "seamless")?
8. State coverage: empty, loading, error, extremes — all present?
9. Is prefers-reduced-motion honoured?
10. Does contrast clear WCAG AA?

Fix what you hit, then tell me what you changed.

The thing tooling can't fix

Three gates catch most of the surface-level tells. What they can't catch is content with nothing in it.

A page can pass every rule, use a professional palette, cover every state, and still say nothing — and no tool helps with that. That's a step-zero failure: you hadn't decided who you were persuading, or of what.

Which is why step 0 of the deck workshop is placed where it is. Every quality mechanism is downstream. Upstream is your own judgement.

Related

• 04 · Craft — the full list of twelve
• Get started / The discovery form — gate one in practice