Agent-built demo / After audit
Margin
AI chat and copilots · Nov 8, 2026
3 measured views. Local app. Design grades compare the actual screen with published product references. N/A means there was no applicable sample. Unverified means evidence was insufficient.
Product specificity and generic signals · A
Both widths identify Field notes, a field guide binding question, grain direction, a 128-page extent, 3 mm bleed and sewn-section testing. The Studio brief supplies a concrete reference; Message Margin identifies the continuation action. (/chat/field-notes)
Visual specificity before deductions A. Weighted design base B. Distinct signal deductions 0.
Product specificity supplies 55% of the base, and the five visual design dimensions share 45%. Round the base to the nearest letter. Each distinct verified generic signal lowers it by one letter, down to F. Foundations are reported separately. The detector ceiling is A. Vision reviewed 1 of 1 measured routes.
- Template copy · not observed
- No instance was identified in the reviewed views.
- Gradient hero · not observed
- No instance was identified in the reviewed views.
- Emoji bullets · not observed
- No instance was identified in the reviewed views.
- Default shadcn grays · not observed
- No instance was identified in the reviewed views.
- Purple on white · not observed
- No instance was identified in the reviewed views.
- Centered everything · not observed
- No instance was identified in the reviewed views.
Compared with the 2026 references
- /chat/field-notes · Like the supplied arena reference, desktop Margin anchors a specific answer with a dark navigation rail. The reference keeps the response decision beside the answer; Margin’s phone layout separates its corresponding next action, replying, with the Studio brief. Open WebUI · evaluating an arena response
Visual design
- Hierarchy
- C
- 1440 keeps the answer, reference and composer in a clear workspace. At 390 the reference interrupts the answer-to-reply sequence and pushes the input and send action below the captured viewport. (/chat/field-notes)
- Typography
- B
- Both widths distinguish the thread title, speaker labels, strong lead advice and supporting checklist. At 390 the bold answer occupies several dense lines, but the checklist remains readable. (/chat/field-notes)
- Spacing
- B
- 1440 uses consistent gutters and separates the reference from the conversation. At 390 cards retain comfortable side margins and padding, although the stacked reference increases the distance to replying. (/chat/field-notes)
- Copy and labels
- A
- Both widths use precise production language and actionable checks. Field notes, Studio brief and Message Margin establish the record, supporting document and next task without promotional filler. (/chat/field-notes)
- Imagery
- C
- 1440 shows the complete book illustration within the Studio brief. At 390 the enlarged illustration is visibly clipped at its lower edge, weakening the reference card’s finish while its caption stays readable. (/chat/field-notes)
Foundations · A
Contrast, responsive layout and keyboard reach are sampled separately. The foundations grade is the lowest verified result in this group.
- Color contrast
- A
- At least four opaque text samples per view passed the contrast threshold. Verified across 3 measured views.
- Responsive layout
- A
- All sampled viewports fit without horizontal overflow and contained readable text. Verified across 3 measured views.
- Keyboard access
- A
- Tab reached two visibly focused controls and Shift+Tab returned to the first. Verified across 3 measured views.
Fix these first
No repair could be verified from these views.
Evidence from the screens
Gray blocks conceal form values or sensitive text. Numbered outlines identify measured issues.
No issue was detected in this view.
No issue was detected in this view.
No issue was detected in this view.
Independent task review
The reviewer assesses each route against its user task using desktop and phone evidence. Its task grades describe the interaction evidence available for that route.
/chat/field-notes
Audit Margin’s working thread at /chat/field-notes: read the field guide binding advice, consult the Studio brief, and send a follow-up message.
Reviewed at 1440 × 900 and 390 × 844.
- Task hierarchy
- C
- 1440: the question, binding advice and message composer are visible together. 390: the answer is readable, but the Studio brief precedes the composer; only its label appears at the bottom, leaving the next conversation action outside the opening view.
- Landing message
- N/A
- This is an internal conversation screen, not a public product landing page.
- Interaction feedback
- Unverified
- Neither supplied check records typing or sending a message. The desktop Send message control appears muted, but its enabled state, submission feedback and recovery behavior cannot be established.
- Access and recovery
- B
- Both checks report reaching a visible Tab focus target. Labels and thread navigation are visible at 1440 and 390, and phone text fits the viewport. Full keyboard reach, composer focus, contrast measurements and error recovery remain unverified.
- Release evidence
- Unverified
- Evidence establishes two local renders, one primary heading at each width, nine visible desktop controls, four phone controls and a reached focus target. Message completion, persistence and deployed behavior were not observed.
First action
At 390, place the message composer before the Studio brief, or collapse the reference card, so continuing the Field notes conversation is easier to find after reading the advice.
How to verify Recheck at 390 × 844: confirm the composer follows the answer, then Tab to its input and Send message control. Submit a follow-up about the folded sheet and observe pending, success and failure feedback. Repeat submission at 1440 × 900.
What a repair could look like
Margin is an agent-built chat workspace demo. Both captures show the same task.
