メニュー
Use the shipped examples like a user in a real browser and fix the interaction defects the sweeps can't see — placement, tracking, legibility, crowding. Run after runtime or theme changes, or as a standing dispatch.
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/max-sixty/colloquy --skill ui-sweepコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
SKILL.md を表示中
| name | ui-sweep |
| description | Use the shipped examples like a user in a real browser and fix the interaction defects the sweeps can't see — placement, tracking, legibility, crowding. Run after runtime or theme changes, or as a standing dispatch. |
The suite holds the invariants somebody has already stated: render_version gates a page's rendering, and the press and poll sweeps hold it still under the aim. The defects that reach the user first are the ones nobody has stated yet — a float that parts from its passage on scroll, a diagram scaled below legibility, an input that stops growing at ten lines. Those are found by using the page. This skill is that use: drive, judge, fix, pin.
Serve pages the way tests/test_render.py does (serve, open_page, or the same
pattern in a scratch script) and drive them with real input. Tour the gallery and
one prose-heavy example, at 1200×900 and 1440×900, in both color schemes. At each
station take a screenshot and record the geometry it claims:
Read the frames against plugins/colloquy/skills/colloquy/CLAUDE.md — each norm there is a checklist row — and against the craft the norms don't state: nothing stands on words being read or written about, nothing renders below legibility, nothing runs out of room while the screen has some, floats track what they point at. A finding is a reproduced number or screenshot, not an impression.
Fix each finding, then pin it where the gate and the suite share it: a fact about a rendered page goes in render_version, a fact about a gesture becomes a test in test_render.py. Put the bug back once and watch the new check fail. A judgment call — removing a control, redesigning a flow — goes in the report with its screenshot, not in the diff. The run ends as a green branch and a report, and landing waits for the go-ahead.