Super
Design knowledge for a build mode that checks its own work.
This review is gated.
Design knowledge for a build mode that checks its own work.
This review is gated.
design knowledge, checked once
Three kinds of evidence, kept apart by how much stands behind them.
119 design categories on four axes (type of site, domain, purpose, visual style) and 75 app patterns such as checkout, booking or chat.
Each rule is labelled by its evidence: standard, research, or site observation.
rules picked for the prompt, used twice
For each user prompt, one category per axis where it applies, plus up to six app patterns. A style is chosen only when the prompt describes a look.
The agent’s planning step does this itself from the skill’s routing table (about 4.6k tokens to read). When the prompt names no look, the planning step can ask one plain clarifying question about the visual style, with a safe recommended default, and fill the plan’s visual identity section from the matched categories.
Packaged as a skill folder in the agent’s own format: a routing table and one reference file per category and app pattern. The agent’s planning step reads the matched files by itself, so nothing is appended by hand and only the matched rules are read: a median of 11.7k tokens per prompt (up to 16.1k).
The same rules also exist as one precompiled instruction per prompt, median 9.9k tokens, for a test that appends it to the system instruction directly.
Prototype instructions are precompiled for 126 user prompts that were not used to build the graph, with two comparison instructions if useful: a short version of about 662 tokens, and one compiled for unrelated categories.
Packaged as a second skill that adds to the verification plan (journeys, screens, checks) and to the critique (request-specific checks filed under its rubric categories, each with a severity). After the build, the same matched categories choose which screens are worth capturing (a filled cart, a form error, a confirmation) and which yes/no questions to ask about each. Code compares the answers with the expected ones and turns failures into fixes.
A first trial on 33 apps showed the questions caught little on first-load screens; they now carry roles and a capture plan. This part is a concept, not yet tried in this form.
unchanged
What the user asked for, in their words. It is the only input the categories are picked from.
The coding agent writes the app. Use 1 adds the design instruction to its system instruction; nothing else changes.
The app runs, so its real states can be reached and seen.
Captures of the running app, at phone and desktop widths. Use 2 would decide which states are captured.
A vision model looks at the captures and code checks measure the page. Use 2 would supply precise yes/no questions instead of an open review.
Failures become concrete fixes for one repair round, then what failed is checked again.
“Implement a chair shop whose dark product photographs use tight spotlights; keep the surrounding page charcoal, with prices and purchase controls.”
2 of the 219 rules in this prompt’s instruction, as written.