Shipyard UX redesign
Part of the Shipyard plugin. If it was installed by itself (for example through skills.sh, ClawHub, Hermes, pi or aiden) and Shipyard's
scripts/folder is not on this computer, install the full plugin first by following https://github.com/philipposk/shipyard/blob/main/INSTALL.md. Steps 1 to 5 need only a browser and Node; step 6 builds in the student's own repo.
You turn a cluttered app into a calm one: an audit from the code, real user workflows with step counts, patterns from similar products, 6 to 12 phone mockups the student can look at, and then small pull requests behind a beta switch so the old layout keeps working. Colours and fonts belong to the design skill; this skill moves, merges and removes things.
Rules (always)
- Read the code before proposing anything. Every claim in the audit names the file it came from.
- Every proposal comes with a step count: current vs proposed, per workflow.
- Show mockups and get a yes before building. Never merge; the student merges.
- A look (theme) changes how things look. A layout changes where things are and what is shown. Keep them as two separate settings (see
references/patterns.md, "Themes vs layout"). - Say which skills and tools you actually ran, and which you skipped and why.
1. Audit from code
Write ux-audit.md in a scratch folder (not the repo). List, with file paths:
- Navigation: every bottom-bar item, sidebar entry, header button, per screen size.
- AI entry points: every place the user can ask an AI (tabs, widgets, floating buttons, live panels). More than one is the most common problem.
- Creation entry points: upload, record, paste link, import, connect. Count how many places each appears.
- Sheets, modals, drawers, and what opens them.
- The busiest screens (live, detail, editor): count visible controls at 390x844.
Read the app's
DESIGN.md/PRODUCT.mdif present.
2. People and workflows
3 to 4 personas, 5 to 6 workflows (for example: record a meeting and get the action items; upload a file; paste a link; ask across everything; find an old item; share). For each: current steps (counted from the code) and proposed steps. Put it in a table; the totals are the headline.
3. Look at similar products
Web research, not memory, for 4 to 6 products that solve the same problem (not only direct competitors: AI chat apps, notes apps, task apps). For each, one pattern worth copying and its source link. Mark anything from memory as such. references/patterns.md has the patterns that held up last time.
4. Principles
Write 5 to 7 principles the mockups must obey, starting from the defaults in references/patterns.md (one AI entry point scoped to context with citations; Home, +, Actions bar plus an avatar drawer; one + sheet with smart defaults; a calm live screen with at most 5 controls; summary-first detail screens; processing screens that are useful while waiting). Drop any that do not fit this app.
5. Mockups
Static HTML, one file per screen, phone frame 390x844, the app's real colours from its CSS tokens. Name them 01-home.html ... Then screenshot each with Playwright and build index.html showing all screens with a one-line reason under each and the workflow table. Script: references/shot.mjs. Look at every PNG yourself before sending.
If the Superdesign CLI is installed and signed in, generate 2 to 3 of the screens there as a second opinion and link the project. If its setup step fails, say so and continue with the HTML.
For ideas, use an ideation skill (for example ce-ideate or brainstorming) if installed; a light single pass is fine when usage is tight. Say which.
Send: the index link or screenshots, the step-count table, and a short "what changes" list. Ask which direction to build.
6. Build behind a beta switch
- First PR: groundwork only (lint/test scripts, shared helpers the later PRs need).
- Add a setting "New layout (beta)", off by default, separate from the theme picker. One attribute on
<html>(for exampledata-layout="v2") set before paint, so there is no flash. - One PR per principle, stacked in order: one AI, navigation, + sheet, live screen, detail screen, cross-item list (Actions), then desktop.
- Each PR: unit tests, type check, build, the app's end-to-end tests (note failures that also fail on main), phone screenshots at 390x844 in the PR body, and what was and was not tested (for example "not tried on a real phone").
- Review each PR before opening it (a code-review skill, or by hand against the diff if the skill's files are missing; say which), then verification-before-completion.
7. Common mistakes
See references/patterns.md, "Pitfalls". Check each before saying a PR is ready.