AGENT SKILL

look

Look at the web app as a person would — take or read the e2e's screenshots of the pages a change touches, in a light and a dark theme, open the pages with the browser tools and photograph them, and name what is cut off or overlapping, unreadable in a theme, or the first thing a person would notice. For a step that changes a page, before it signs off, and for one that reviews the change. Writes nothing into the workspace.

License 许可
MIT
所属仓库
emyann/nemia
最近更新
2026-11-15
Stars
★ 3
Forks
0

01 前言 Front Matter

name
look
description
Look at the web app as a person would — take or read the e2e's screenshots of the pages a change touches, in a light and a dark theme, open the pages with the browser tools and photograph them, and name what is cut off or overlapping, unreadable in a theme, or the first thing a person would notice. For a step that changes a page, before it signs off, and for one that reviews the change. Writes nothing into the workspace.

03 SKILL.md 正文

Look at the web app as a person would

A page that passes its assertions and looks wrong is not done. This is how a step looks at the web app (apps/web) the way a person who opens it would: through pictures of its pages in both themes, and through the pages themselves in a browser. Any step may use it: the one that changes a page, before it signs off, and the one that reviews the change, beside the diff.

Write nothing into the workspace

A file written into the worktree is committed by the run and rides the branch into what ships. Every picture goes outside it.

  • Yours go in your files directory: $WF_ARTIFACTS when it is set, or the directory your brief names under Files for the steps after you, which is the one the run gives your check as $WF_ARTIFACTS. With neither, a directory of your own under the system's temp directory (node -p "require('os').tmpdir()").
  • The browser saves only where its server lets it: its output directory (--output-dir in .mcp.json), or the workspace, which is never where you save. In a run, its output directory is your step's files directory, $WF_STEP_FILES, kept with the run; outside one, /tmp/playwright-mcp, shared by every session on the machine. Give browser_take_screenshot an absolute filename under that directory ($WF_STEP_FILES/<page>-<theme>.png). A relative filename is resolved against the workspace, where you never save, and an absolute path anywhere else is refused (File access denied … outside allowed roots). (Measured with Claude Code 2.1.296 and @playwright/mcp 0.0.80.)
  • A step that only reads gives no filename at all. A design or a review step is refused one on every browser tool that takes it (browser_take_screenshot, browser_snapshot, browser_console_messages, browser_network_requests, browser_network_request, browser_evaluate), since a file named could be the workspace's. Called without one, the tool saves in its output directory, says where, and answers with the picture itself: judge that. In a run it says a path relative to your workspace that lands in $WF_STEP_FILES: name the picture you show a person by its absolute path there, $WF_STEP_FILES/<file> written out, in artifacts and in the question that asks about it, never a /tmp path.
  • The e2e too: run it with WF_ARTIFACTS set. Without it, it writes its pictures into .e2e/ at the repository root.

1. Take the pictures, or read them

npm run e2e builds the web app, serves it against a seeded store, and visits every page in apps/web/test/e2e/pages/ in headless Chromium, in the dark theme, in the order of their file names. It takes no picture but a failing page's, <page>-<theme>-failed.png, and then exits non-zero.

npm run e2e -- --photos --pages <name,…> takes them: it visits in a light and then a dark theme, and pictures the pages named, each as <page>-<theme>.png with the pictures the page takes of itself (<page>-<what>-<theme>.png, its phone's among them where it sets one), into $WF_ARTIFACTS, and visits no other: a page reads only the seed and what it writes itself. A page is named as its file is, without its number (pages/NAMES lists them); --photos alone pictures every page. Its last lines list the pictures.

--pages runs none of the look's browser tests (apps/web/test/e2e/*.test.ts: the parts on their stories page, the bars, the checks). A change to a part of the look (src/components/look) or a story runs them too: npm run e2e -- --look, or npm run e2e -- --look --pages <name,…> when pages are due as well.

  • When a step before you named pictures in its artifacts or screenshots, those absolute paths are the pictures: read them.
  • Otherwise, take them: WF_ARTIFACTS=<your files directory> npm run e2e -- --photos --pages <name,…> on the change as it stands, again after each change you make to a page.
  • A state no page reaches is pictured in the browser (section 2), or by a page that reaches it.

Read every one with your Read tool, in both themes. Each is the whole page, top to bottom, at 1280 pixels wide: a menu fixed to the window is drawn only as far down as the first screen, which is the capture and not the page.

2. Open the pages in the browser

The pictures are what the e2e saw. The browser is where you see what a person does with a page. Serve this workspace's board with one command, bare, from the workspace's root:

  1. npm run board:serve. It builds the app outside the workspace, serves it on ports of its own against a store of its own, and prints board on http://127.0.0.1:<port> and api on http://127.0.0.1:<port>. Open the first. A board already served from this workspace is said again, and nothing is started.
  2. npm run board:stop when you are done, and always before you sign off. A board nobody stops ends by itself after thirty minutes.

The served store holds the e2e's seed (apps/web/test/e2e/seed.ts): the runs, the session and the loop its pictures show are on these pages, and what you press there writes to that store alone, which goes with the board. It was built when you served it: after a change to a page, stop it and serve again.

Never the live board. :8787 and :5173 are a person's own board over the default store, where a press starts a run or approves a gate. A design or a review step is refused both by address; no step opens them.

Then, for every page you are looking at, in the browser tools (browser_*, from .mcp.json):

  • browser_resize to 1280 × 800, the e2e's window, once.
  • browser_navigate to the page. browser_snapshot says what is on it and what can be pressed, and browser_console_messages what went wrong.
  • browser_take_screenshot with fullPage: true: with an absolute filename under $WF_STEP_FILES (in a run; /tmp/playwright-mcp outside one), then read the file; or, in a design or a review step, with no filename, and judge the picture the call answers with.
  • Use the page as a person would — browser_click, browser_hover, browser_type, browser_press_key on what the change is about — and photograph what it shows.

Both themes. The app follows the system's colour scheme (prefers-color-scheme), and so does its canvas (React Flow's colorMode="system"). The browser's scheme is light, and no browser tool changes it: photograph the light theme there, and take the dark from the e2e's <page>-dark.png under --photos, side by side with the light. Do not restyle a page from browser_evaluate to fake the dark: the canvas would stay light, and you would be judging your own stylesheet.

Never browser_run_code_unsafe. It runs code of your choosing in the browser server's own process, past every fence the other tools keep. Every shipped step that could reach it is denied it by name, in its agent's disallowedTools:; a job in dontAsk is refused it because no allow reaches it.

When the browser is denied. A step in permissionMode: dontAsk runs only what an allow reaches: its agent's allowedTools:, or .claude/settings.json. One whose agent allows no browser_* tool, or that names no playwright server, has no browser, and is refused cp too; reading a PNG outside the workspace is allowed. Then the pictures are the pages: read them, say in what you report that you judged from the screenshots, and do not guess at what a page does when pressed.

3. Judge what you see

Judge the pages against what was asked for:

  • Layout. Something cut off, overlapping, misaligned, or pushed off the screen; a page empty where it should not be; text that runs out of its box.
  • Both themes. Text that cannot be read; a colour that belongs to the other theme; a border, a background or an icon that vanishes in one of them.
  • What a person would notice first. Whether the page does what was asked, and anything that looks broken or unfinished.

Each thing that is off names the page and the theme, the screenshot it shows in, and what is wrong and where on the page. Read the code only to understand what a page is meant to show. A screenshot that is missing or cannot be read is a finding. If you find nothing a person would notice, say so: a look that invents work to look thorough is worse than a short one.

If the change is yours, fix what is off and look again before you sign off.

04 同仓库更多 Skills

emyann
emyann/nemia

groom-the-inbox

Read each card in play in the inbox against the repository and say, per card, done (the commit on the base that landed it), replaced (the decision or card that did), stale (each fact that moved, what it said and what it says now) or standing, reading closely only what the period touched. For the janitor's step, which proposes and never acts: a person removes or amends each card by the command the check writes.

★ 3
emyann
emyann/nemia

write-a-spec

Turn an accepted intent into a spec an engineer can build from in this repository: the goal restated, the approach and why, every path it touches, the steps, the tests that prove it, the risks, and what is still open. For a step that designs a change before anyone builds it.

★ 3
emyann
emyann/nemia

review-a-change

Review a change in this repository as a second engineer would: against its intent and spec, by reading the diff and the suite's verdict, with findings that name the file, the line and what goes wrong. For a read-only step that decides whether the work goes round again.

★ 3
emyann
emyann/nemia

read-the-stats

Read a factory's stats since the janitor's last run, against the same length of time before it, write a short report, and propose one correction per card — an allow or deny by name, a line in an agent's body or a skill, or a refusal kept on purpose — each with the numbers behind it, never one already in the inbox. For the janitor's step, which proposes and never acts.

★ 3
emyann
emyann/nemia

land-a-branch

Land a run's result on the base branch in this repository's own checkout, the way the step says it lands: make that one act, read its result, and write the message that records what landed — or change nothing and say why it refused. For a gated ending step (workspace: base).

★ 3
emyann
emyann/nemia

implement-a-change

Build a change from its spec in this repository: tests first where a rule is claimed, small commits, the tests of what changed green for the right reason, and each review finding answered in the next round. For a step that writes the code.

★ 3