review-usability

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Website Usability Review

网站可用性审查

Evaluate how effectively people can use a running website: can they learn it on first visit, finish tasks quickly, find their way back later, avoid mistakes, and leave satisfied. This skill judges rendered pages and real interactions, not source code.
Scope against
review-accessibility
: completing a top task by keyboard is in scope here, because the article treats keyboard access as efficiency (Mistake 4). Screen-reader behavior, ARIA, color contrast, and WCAG conformance are not; note them in one line and point to
review-accessibility
.
评估人们能多有效地使用一个正在运行的网站:他们能否在首次访问时学会使用它、快速完成任务、之后重新找到路径、避免错误,并带着满意感离开。本 skill 评判的是渲染后的页面和真实交互,而不是源代码。
与
review-accessibility
的范围划分:通过键盘完成一项主要任务属于本 skill 的范围,因为文章将其视为效率问题(错误 4)。屏幕阅读器行为、ARIA、色彩对比度和 WCAG 一致性不在此范围内;请用一行说明,并指向
review-accessibility
。
参考:软件可用性基础——另见 Fundamentals

Domain Knowledge

领域知识

The rubric comes from the article itself, pulled fresh on every run so the review tracks the current version of the article rather than a copy frozen into this skill.
Download it before doing anything else:
bash
ARTICLE_URL="https://jeffbailey.us/blog/2026/01/01/fundamentals-of-software-usability/llm.txt"
CACHE_DIR="${TMPDIR:-/tmp}/review-usability"
ARTICLE="$CACHE_DIR/fundamentals-of-software-usability.txt"
mkdir -p "$CACHE_DIR"
if curl -fsSL --max-time 20 -A "skills-review-usability/1.0" "$ARTICLE_URL" -o "$ARTICLE.tmp" \
   && grep -q "Learnability" "$ARTICLE.tmp"; then
  mv "$ARTICLE.tmp" "$ARTICLE"
  echo "SOURCE: live $ARTICLE_URL (fetched $(date '+%Y-%m-%d %H:%M %Z'))"
else
  rm -f "$ARTICLE.tmp"
  echo "SOURCE: fallback references/wisdom.md (live fetch failed)"
fi
Then pull out only the sections the review uses. They are about a quarter of the article (roughly 11 KB of 41 KB), and the rest (introductions, misconceptions, future trends) never changes a score:
bash
SRC="$ARTICLE"; [ -f "$SRC" ] || SRC="${CLAUDE_SKILL_DIR:-skills/review-usability}/references/wisdom.md"
awk '/^## /{k2=($0 ~ /^## Section (6|7|9):/); k3=0}
     /^### /{k3=($0 ~ /^### (When .* Fails|Quick Check:)/)}
     k2||k3' "$SRC" > "$CACHE_DIR/rubric-sections.md"
grep -c '^#' "$CACHE_DIR/rubric-sections.md"   # expect about 24 headings
Read
rubric-sections.md
. If it has fewer than about 15 headings, the article has been restructured: read the source in full instead. When the live fetch failed, the source is
references/wisdom.md
, the same article synced weekly by
.github/workflows/sync-wisdom.yml
; its
Last updated:
line dates the snapshot.
Record the
SOURCE:
line in the report header, so a reader knows which version of the rubric produced the scores.
What each extracted section is for:
  • When … Fails and Quick Check for each of the five dimensions: the questions each dimension is scored against. Open a dimension's Design Patterns subsection only when a remediation needs it.
  • Common Usability Mistakes (Section 7). Treat each one as a named finding type.
  • Usability Testing (Section 6). It shapes the "Validate with real users" section of the report.
  • When NOT to Focus on Usability (Section 9). Use it to keep recommendations proportionate to the site.
Browser scripts for the walkthrough (viewport size, content offset, keyboard reach, shortcut probes, render-error scan, Back-button probe) live in
references/browser-recipes.md
. Use them instead of writing new ones each run.
If the article's content contradicts this file, the article wins. It is the source of truth; this file only describes the procedure.
评分标准直接来自文章本身;每次运行都会重新拉取,使审查依据的是文章的当前版本,而不是冻结在 skill 内的副本。
在开始其他操作前下载:
bash
ARTICLE_URL="https://jeffbailey.us/blog/2026/01/01/fundamentals-of-software-usability/llm.txt"
CACHE_DIR="${TMPDIR:-/tmp}/review-usability"
ARTICLE="$CACHE_DIR/fundamentals-of-software-usability.txt"
mkdir -p "$CACHE_DIR"
if curl -fsSL --max-time 20 -A "skills-review-usability/1.0" "$ARTICLE_URL" -o "$ARTICLE.tmp" \
   && grep -q "Learnability" "$ARTICLE.tmp"; then
  mv "$ARTICLE.tmp" "$ARTICLE"
  echo "SOURCE: live $ARTICLE_URL (fetched $(date '+%Y-%m-%d %H:%M %Z'))"
else
  rm -f "$ARTICLE.tmp"
  echo "SOURCE: fallback references/wisdom.md (live fetch failed)"
fi
然后仅提取审查会用到的章节。这些内容约占文章的四分之一(大约 11 KB/41 KB),其余部分(引言、常见误区、未来趋势)不会影响评分:
bash
SRC="$ARTICLE"; [ -f "$SRC" ] || SRC="${CLAUDE_SKILL_DIR:-skills/review-usability}/references/wisdom.md"
awk '/^## /{k2=($0 ~ /^## Section (6|7|9):/); k3=0}
     /^### /{k3=($0 ~ /^### (When .* Fails|Quick Check:)/)}
     k2||k3' "$SRC" > "$CACHE_DIR/rubric-sections.md"
grep -c '^#' "$CACHE_DIR/rubric-sections.md"   # expect about 24 headings
阅读
rubric-sections.md
。如果标题少于约 15 个,说明文章结构已经调整:改为阅读完整源文。实时拉取失败时,源文件为
references/wisdom.md
,即由
.github/workflows/sync-wisdom.yml
每周同步的同一篇文章;其
Last updated:
行注明快照日期。
在报告头中记录
SOURCE:
行,让读者知道用于评分的评分标准版本。
提取的每个章节有何用途:
  • 每个维度的 When … Fails 和 Quick Check:用于确定各维度的评分问题。仅在制定修复方案需要时,才打开某个维度的 Design Patterns 子章节。
  • Common Usability Mistakes(第 7 节):将其中每一项视为一种具名发现类型。
  • Usability Testing(第 6 节):用于构建报告中的“使用真实用户验证”部分。
  • When NOT to Focus on Usability(第 9 节):用于确保建议与网站规模相称。
用于走查的浏览器脚本,包括视口大小、内容偏移、键盘可达性、快捷键探测、渲染错误扫描和返回按钮探测,位于
references/browser-recipes.md
。应使用这些脚本,而不是每次运行时重新编写。
如果文章内容与本文件存在冲突,以文章为准。文章是事实来源;本文件仅描述操作流程。

Workflow

工作流

  1. Load the rubric. Run the fetch and extraction above and read the extracted sections.
  2. Profile the site. Take the target URL from the user. If none is given, ask for one. Identify what kind of site it is (blog or docs, marketing, e-commerce, web app, or a single static or placeholder page) and its primary visitors. If the user has analytics or Search Console data, use it to find the pages and entry points real visitors use. Otherwise, infer them from the navigation, the sitemap (
    /sitemap.xml
    ), and the home page.
    • Local target: a local build or preview server is fine for checking fixes before they deploy. Some things only the real host shows: HTTP caching headers, the custom 404 page (a static server substitutes its own), compression, and speed. Check those against the live URL, and label each finding with the host it came from.
    • Re-review: if a previous usability report exists (the user names it, or it sits at the output path), read it first. Walk the same top tasks so the scores are comparable. The report then adds a "Prev" score column and a "Prior Findings" table giving each earlier finding's status (Fixed, Still present, Deferred, or Not verifiable here) with one line of evidence. Also probe the fixes for regressions: edge-case input (special characters in search, such as
      c++
      and
      a&b
      ), states that should stay hidden, and links the fix added. Check each fix from both sides: that the problem is gone, and that whatever sits next to it still works (removing one search tag must not hide titles or excerpts).
  3. Define 1 to 5 top tasks, as many as the site really supports. These are the things a visitor came to do. For a content site that usually means: land on an article from search and get the answer; find related content; search the site; browse a topic or category; copy a code sample; subscribe or follow. A single static page may support only one or two (read it, follow its link); don't pad the list with tasks the site has no reason to offer.
    Confirm the task list with the user when they are available, because the wrong tasks make the whole review wrong (the article's "When Usability Testing Fails"). When nobody can confirm (a non-interactive run, or the user said "just do it"), infer the tasks from the navigation, sitemap, and home page, proceed, and mark each one Unconfirmed (inferred from <source>) in the Top Tasks table so the reader knows to check them.
    Budget. A full checklist pass on every task at two viewports does not fit one session, so set the scope before walking: at most about 12 distinct pages, every task at desktop, the one or two most important tasks also at phone width and by keyboard, and the high-yield checks first (answer placement, search, related content, Back, 404, phone layout). Anything the budget leaves out goes under Not Verified by name, so silence is never read as a pass.
  4. Walk each task. Use a browser automation tool if one is available (Playwright, Chrome DevTools, or a similar MCP server), at a desktop viewport (about 1280px) and, for the tasks the budget names, a phone viewport (about 390px) and keyboard alone. For each step, record the URL, what was clicked or typed, what happened, and a screenshot path when you can capture one. With no browser tool, fall back to fetching HTML with
    curl
    and mark every interaction-dependent check (search results, menus, focus behavior, animations) as not verified rather than guessing.
    • Recipes: use the scripts in
      references/browser-recipes.md
      for viewport size, content offset, keyboard reach, shortcuts, render errors, and the Back probe.
    • Viewports: if the tool cannot set an exact width (a maximized window often refuses to resize), use device emulation or open a popup at the target size, for example
      window.open(url, "phone", "popup,width=390,height=844")
      from a page script, then switch the tool to the new page. Browser chrome makes the inner height smaller than requested. Any width within about 50px is fine. Record the actual width and inner height of each viewport in the report header: the desktop "first screen" check depends on height, and a maximized window with browser chrome is often only 600-700px tall.
    • Cache: start from a fresh profile or reload with the cache bypassed, so you judge what a new visitor gets. Then check what a returning visitor gets: read the HTML response's
      Cache-Control
      with
      curl -sI <url>
      . A long
      max-age
      on HTML means returning visitors can see stale pages (a memorability finding). A missing
      Cache-Control
      header is not a finding by itself; note it in one line, since browsers then cache heuristically from
      Last-Modified
      .
    • Reading page state: prefer small scripts that return compact JSON (element offsets, computed styles,
      document.activeElement
      , result counts) over full accessibility snapshots. On long pages a snapshot can exceed 60,000 characters per action, and some tools (Chrome DevTools
      press_key
      ) attach one to every key press, even a single Tab. Check keyboard reach from the page's tab order, and fire shortcuts by dispatching a
      KeyboardEvent
      from a script (recipes in
      references/browser-recipes.md
      ).
    • Shortcut probes may navigate. A
      /
      or
      Ctrl+K
      handler can load a search page, which destroys the script context ("Execution context was destroyed"). That is a result: read
      location.href
      in a new call and record where the shortcut went.
    • Render errors: scan only text outside
      pre
      and
      code
      , because articles often show error messages in code samples.
    • Timing: after an action that triggers a CSS transition, wait past it before reading computed styles, or you will read the starting value.
    • Back button: browsers often restore a page from the back/forward cache without running any of its scripts, so a Back test can pass even when the page's own restore code is broken. Set a marker first (
      window.__probe = 1
      ); if the marker survives Back, the cache served the page. Then reload to exercise the restore code, and report both results.
  5. Run the checklist. Evaluate the walked pages against
    references/checklist.md
    . Every finding needs a URL plus the element or step where it occurred.
  6. Translate before judging. The article's patterns are written for applications in general. Map each one to this site before scoring it, for example: "undo" on a content site is a working back button and reversible filters; "bulk operations" rarely apply. Mark a check N/A when the site has no matching interaction, and never lower a score for a pattern the site has no reason to offer.
    • "Answer in the first screen": measure where the main content starts, in pixels and in screens. On a 390px-wide phone, content that starts beyond about 1.5 screens (roughly 1,250px) is a finding. Navigation that is the route to the answer, such as an open list of fixes on an error-fix guide, counts as content, not as clutter.
    • Empty search results: search for a nonsense string (for example
      qzxwvkjh
      ). If fuzzy matching still returns results, try one more string, then mark the check not verified instead of passing it.
    • Page speed: measure with a DevTools performance trace or Lighthouse against the real host. Local static servers (such as
      python -m http.server
      ) have no compression or CDN and give misleading numbers. When a page is slow only in the browser session and fast with
      curl
      , list it under Not Verified.
  7. Score each dimension on a 1-10 scale against its quick-check questions. A score needs evidence from the walkthrough. When a dimension has fewer than two checks the site gives you anything to observe (on a single static page, Error Prevention may have only a 404), write N/A — <reason> instead of a number, or give the score with "(thin evidence: <what was observed>)". The Overall score is the mean of the numeric dimensions, one decimal, and the report says which were averaged.
  8. Find patterns. An issue repeated across a page template (every article, every category page) outranks a one-off, because fixing the template fixes every page.
  9. Self-check, then write the report (format below), including a short plan for testing the top tasks with real users. Before writing, confirm:
    • every Detailed Finding's Location has a full
      https://
      URL (or the local URL plus host label) and the element or task step; a finding that only names a template ("article template at 390px") gets the URL of the page where it was seen;
    • every finding has confidence >= 7; drop the rest or move them to Not Verified;
    • the header records both viewports as width x height and the browser tool;
    • each top task is marked confirmed or Unconfirmed, and there are 1 to 5 of them;
    • the Overall equals the mean of the numeric scores.
  1. 加载评分标准。 执行上述拉取和提取操作,然后阅读提取出的章节。
  2. 分析网站。 从用户处获取目标 URL。如果未提供,则询问。识别网站类型(博客或文档、营销、电子商务、Web 应用,或单个静态/占位页面)及其主要访客。如果用户提供分析数据或 Search Console 数据,使用它来查找真实访客使用的页面和入口。否则,根据导航、站点地图(
    /sitemap.xml
    )和首页进行推断。
    • 本地目标: 本地构建或预览服务器可用于在部署前检查修复效果。有些问题只有真实 host 才能显示,例如 HTTP 缓存标头、自定义 404 页面(静态服务器会替换成自己的页面)、压缩和速度。针对实时 URL 检查这些问题,并在每条发现中标明它来自哪个 host。
    • 复审: 如果之前存在可用性报告(用户指定了报告,或报告位于输出路径),先阅读该报告。走查相同的主要任务,使分数具有可比性。然后在报告中增加“Prev”分数列和“Prior Findings”表格,通过一行证据说明每项旧发现的 status(Fixed、Still present、Deferred 或 Not verifiable here)。同时探测修复是否引入回归:检查边界输入,例如搜索中的特殊字符
      c++
      和
      a&b
      、本应保持隐藏的状态,以及修复新增的链接。从两个方面检查每项修复:确认问题已经消失,并确认相邻功能仍然可用(移除一个搜索标签不应隐藏标题或摘要)。
  3. 根据网站实际支持的数量定义 1 到 5 个主要任务。 这些任务是访客来到网站要完成的事情。对于内容型网站,通常包括:从搜索结果进入一篇文章并找到答案;查找相关内容;站内搜索;浏览某个主题或分类;复制代码示例;订阅或关注。单个静态页面可能只支持一两个任务(阅读、访问其链接);不要用网站没有理由提供的任务凑数。
    如果可以联系到用户,应与用户确认任务列表,因为错误的任务会让整个审查失去意义(文章中的“When Usability Testing Fails”)。如果无法确认(非交互式运行,或用户表示“直接做”),则根据导航、站点地图和首页推断任务并继续执行,同时在 Top Tasks 表中将每项标记为 Unconfirmed (inferred from <source>),让读者知道需要自行核实。
    预算。 在一次会话中,无法对每项任务在两个视口下执行完整检查,因此走查前应确定范围:最多检查约 12 个不同页面;每项任务均检查桌面端;最重要的一到两项任务还需检查手机宽度和纯键盘操作;优先执行收益较高的检查(答案位置、搜索、相关内容、返回、404、手机布局)。预算未覆盖的内容应按名称列入 Not Verified,避免将沉默误认为通过。
  4. 走查每项任务。 如果有可用的浏览器自动化工具(Playwright、Chrome DevTools 或类似 MCP server),在桌面视口(约 1280px)下执行;对于预算指定的任务,还需在手机视口(约 390px)和纯键盘条件下执行。每一步记录 URL、点击或输入的内容、发生的结果,以及可捕获时的截图路径。没有浏览器工具时,退回到使用
    curl
    获取 HTML,并将每项依赖交互的检查(搜索结果、菜单、焦点行为、动画)标为 not verified,而不是进行猜测。
    • 配方: 使用
      references/browser-recipes.md
      中的脚本处理视口大小、内容偏移、键盘可达性、快捷键、渲染错误和返回探测。
    • 视口: 如果工具无法设置精确宽度(最大化窗口通常会拒绝调整大小),使用设备模拟,或从页面脚本打开目标尺寸的弹出窗口,例如
      window.open(url, "phone", "popup,width=390,height=844")
      ,然后将工具切换到新页面。浏览器 chrome 会使内部高度小于请求值。宽度在约 50px 范围内即可。在报告头中记录每个视口的实际宽度和内部高度:桌面端“首屏”检查依赖高度,而最大化窗口加上浏览器 chrome 后通常只有 600–700px 高。
    • 缓存: 从全新配置文件开始,或在绕过缓存的情况下重新加载,从而判断新访客看到的内容。随后检查回访访客看到的内容:使用
      curl -sI <url>
      读取 HTML 响应中的
      Cache-Control
      。HTML 上过长的
      max-age
      意味着回访访客可能看到陈旧页面(属于记忆性问题)。缺少
      Cache-Control
      本身不构成发现;只需用一行说明,因为浏览器随后会根据
      Last-Modified
      进行启发式缓存。
    • 读取页面状态: 优先使用返回紧凑 JSON 的小型脚本(元素偏移量、计算样式、
      document.activeElement
      、结果数量),而不是完整的可访问性快照。在长页面上,一次快照操作可能超过 60,000 个字符,而且某些工具(如 Chrome DevTools
      press_key
      )甚至会给每次按键都附加一个快照,单次 Tab 也一样。根据页面的 tab order 检查键盘可达性,并通过脚本调度
      KeyboardEvent
      来触发快捷键(配方见
      references/browser-recipes.md
      )。
    • 快捷键探测可能触发导航。
      /
      或
      Ctrl+K
      处理器可能加载搜索页面,从而销毁脚本上下文(“Execution context was destroyed”)。这本身就是结果:在新的调用中读取
      location.href
      ,并记录快捷键跳转到的位置。
    • 渲染错误: 仅扫描
      pre
      和
      code
      之外的文本,因为文章经常在代码示例中显示错误消息。
    • 计时: 触发 CSS transition 的操作完成后,应等待过渡结束再读取计算样式,否则读到的仍是起始值。
    • 返回按钮: 浏览器经常从 back/forward cache 恢复页面而不运行其脚本,因此即使页面自身的恢复代码已经损坏,Back 测试也可能通过。先设置标记(
      window.__probe = 1
      );如果标记在 Back 后仍然存在,说明页面由缓存提供。然后重新加载以执行恢复代码,并同时报告两种结果。
  5. 执行检查清单。 根据
    references/checklist.md
    评估走查过的页面。每条发现都需要提供 URL,以及发生问题的元素或步骤。
  6. 判断前先进行场景映射。 文章中的模式面向一般应用编写。评分前,应将每种模式映射到本站,例如:内容型网站中的“undo”可以是可用的返回按钮和可撤销筛选;“批量操作”则很少适用。当网站没有对应交互时,将检查标为 N/A,且绝不因网站没有理由提供的模式而降低评分。
    • “Answer in the first screen”: 测量主要内容在像素和屏幕数上的起始位置。在 390px 宽的手机上,如果内容起点超过约 1.5 屏(大约 1,250px),则构成发现。导航本身如果是通往答案的路径,例如错误修复指南中展开的修复列表,应算作内容,而不是杂乱元素。
    • 空搜索结果: 搜索一个无意义的字符串,例如
      qzxwvkjh
      。如果模糊匹配仍然返回结果,再尝试另一个字符串,然后将检查标为 not verified,而不是判定通过。
    • 页面速度: 针对真实 host,使用 DevTools performance trace 或 Lighthouse 进行测量。本地静态服务器(例如
      python -m http.server
      )没有压缩或 CDN,提供的数字会产生误导。如果页面仅在浏览器会话中缓慢,而使用
      curl
      时速度很快,则将其列入 Not Verified。
  7. 依据各维度对应的 quick-check 问题和 When … Fails 迹象,以 1–10 分制评分。 每个分数都需要走查证据。当某维度可观察的检查少于两项时(在单个静态页面上,错误预防可能只有一个 404),应写入 N/A — <reason>,而不是数字;也可以给出分数并附上“证据较少:<what was observed>”。Overall 分数是所有数值维度的平均值,保留一位小数,并在报告中说明平均了哪些维度。
  8. 寻找模式。 跨页面模板反复出现的问题(例如每篇文章、每个分类页面都有问题)优先于一次性问题,因为修复模板即可修复所有相关页面。
  9. 自检后撰写报告(格式见下文),并包含一份使用真实用户测试主要任务的简短计划。撰写前确认:
    • 每条 Detailed Finding 的 Location 都包含完整的
      https://
      URL(或本地 URL 加 host 标签)以及元素或任务步骤;仅提及模板的发现(例如“390px 下的文章模板”)必须附上实际看到该问题的页面 URL;
    • 每条发现的置信度都 >= 7;删除其余发现,或将其移至 Not Verified;
    • 报告头记录了两个视口的宽度 x 高度以及浏览器工具;
    • 每项主要任务均标记为 confirmed 或 Unconfirmed,且任务数量为 1 到 5 项;
    • Overall 等于所有数值分数的平均值。

Confidence and Severity

置信度与严重程度

Only report findings with confidence >= 7/10. For each finding, check:
  • Did you observe it on the live site, rather than infer it from how sites usually work?
  • Can you name the URL and the element or step?
  • Would a typical visitor on one of the top tasks actually run into it?
If any answer is no, leave it out. An expert walkthrough predicts problems; only real users confirm them. Say so in the report instead of inflating confidence.
Severity:
  • CRITICAL — a top task cannot be completed (a broken search, a dead primary navigation link, content hidden on mobile).
  • HIGH — a top task completes, but most visitors will struggle or give up (the answer sits below several screens of preamble; there is no route from an article to related content).
  • MEDIUM — friction that slows visitors or makes them hesitate (inconsistent labels across templates, weak link affordance, no search shortcut).
  • LOW — polish (micro-interactions, minor visual inconsistency).
只报告置信度 >= 7/10 的发现。对每条发现检查:
  • 是否在实时网站上观察到了它,而不是根据网站通常的工作方式推断?
  • 能否指出具体 URL 以及元素或步骤?
  • 执行主要任务的典型访客是否确实会遇到它?
如果任何问题的答案是否定的,则不要报告。专家走查可以预测问题,但只有真实用户才能确认问题。应在报告中说明这一点,而不是虚高置信度。
严重程度:
  • CRITICAL——主要任务无法完成(搜索损坏、主要导航链接失效、移动端隐藏内容)。
  • HIGH——主要任务可以完成,但大多数访客会陷入困难或放弃(答案位于数屏前言之后;文章与相关内容之间没有任何路径)。
  • MEDIUM——会减慢访客或使其犹豫的摩擦(不同模板间标签不一致、链接提示不明显、没有搜索快捷键)。
  • LOW——打磨问题(微交互、轻微视觉不一致)。

Scoring Dimensions (1-10 each)

评分维度(各项 1–10 分)

Each dimension is scored against the article's matching Quick Check questions and When … Fails signs:
  1. Learnability — Can a first-time visitor complete the top tasks without help? Familiar layout and icon conventions, discoverable features, clear link and button affordances.
  2. Efficiency — How many steps and how much scrolling does a top task take? Answer placement, search quality and speed, keyboard access, navigation depth, page speed as the visitor feels it.
  3. Memorability — Would a returning visitor find their way again? Consistent navigation, terminology, and visual layout across page templates; stable, predictable URLs.
  4. Error Prevention and Recovery — Are dead ends and mistakes prevented, and easy to recover from? Helpful 404 pages, empty-search results, form validation, a back button that works, reversible filters.
  5. Satisfaction — Does the site feel polished, respectful, and responsive? No layout shift or intrusive interruptions, readable typography, responsive interactions.
每个维度都依据文章中对应的 Quick Check 问题和 When … Fails 迹象评分:
  1. Learnability——首次访问的访客能否在无人帮助的情况下完成主要任务?布局和图标约定是否熟悉、功能是否容易发现、链接和按钮提示是否清晰?
  2. Efficiency——完成主要任务需要多少步骤和滚动?答案位置、搜索质量与速度、键盘访问、导航深度,以及访客感知到的页面速度如何?
  3. Memorability——回访访客能否再次找到使用路径?不同页面模板之间的导航、术语和视觉布局是否一致?URL 是否稳定、可预测?
  4. Error Prevention and Recovery——能否防止死路和错误,并让用户轻松恢复?是否有有帮助的 404 页面、空搜索结果、表单验证、可用的返回按钮和可撤销筛选?
  5. Satisfaction——网站是否感觉精致、尊重用户且响应迅速?是否没有布局偏移或侵扰性中断?排版是否易读?交互是否具有响应性?

Output Format

输出格式

Write the report to the path the user gives. Otherwise, write it to
docs/usability-review.md
in the current working directory, unless that directory is the reviewed site's own repository and the user hasn't asked for the report to live there. In that case, ask for a path or use a scratch or temporary directory, so the review doesn't leave an untracked file in their site. Keep screenshots next to the report.
markdown
undefined
将报告写入用户指定的路径。否则,将报告写入当前工作目录中的
docs/usability-review.md
,除非当前目录是所审查网站自身的仓库,且用户未要求将报告放在该目录。在这种情况下,应询问报告路径,或使用临时/暂存目录,避免审查在网站中留下未跟踪文件。将截图放在报告旁边。
markdown
undefined

Usability Review: <site>

Usability Review: <site>

Rubric source: <the fetch step's SOURCE line, without its "SOURCE:" prefix> Reviewed: <date> · Viewports: desktop <w>x<h>px, phone <w>x<h>px · Browser tool: <name, or "none (HTML only)"> · Target: <live URL or local build>
Rubric source: <the fetch step's SOURCE line, without its "SOURCE:" prefix> Reviewed: <date> · Viewports: desktop <w>x<h>px, phone <w>x<h>px · Browser tool: <name, or "none (HTML only)"> · Target: <live URL or local build>

Summary

Summary

DimensionScoreKey Finding
LearnabilityX/10...
EfficiencyX/10...
MemorabilityX/10...
Error Prevention and RecoveryX/10...
SatisfactionX/10...
OverallX/10Mean of <dimensions averaged>
A dimension may read
N/A — reason
. On a re-review, add a "Prev" column after Score.
DimensionScoreKey Finding
LearnabilityX/10...
EfficiencyX/10...
MemorabilityX/10...
Error Prevention and RecoveryX/10...
SatisfactionX/10...
OverallX/10Mean of <dimensions averaged>
A dimension may read
N/A — reason
. On a re-review, add a "Prev" column after Score.

Prior Findings (re-review only)

Prior Findings (re-review only)

#FindingStatusEvidence
#FindingStatusEvidence

Top Tasks Walked

Top Tasks Walked

TaskConfirmed?Completed?StepsWhere it got hard
TaskConfirmed?Completed?StepsWhere it got hard

Detailed Findings

Detailed Findings

Finding 1: [Title]

Finding 1: [Title]

  • Severity: CRITICAL / HIGH / MEDIUM / LOW
  • Confidence: X/10
  • Dimension: [scoring dimension]
  • Article mistake: [one of Section 7's mistakes, or "none"]
  • Location: full URL (https://...), plus the element or task step
  • Evidence: What you did and what happened, with a screenshot path if captured.
  • Impact: Which visitors are affected, and on which task.
  • Remediation: A concrete change, naming the template or component to edit when it can be identified.
(repeat, ordered by severity; template-wide issues first)
  • Severity: CRITICAL / HIGH / MEDIUM / LOW
  • Confidence: X/10
  • Dimension: [scoring dimension]
  • Article mistake: [one of Section 7's mistakes, or "none"]
  • Location: full URL (https://...), plus the element or task step
  • Evidence: What you did and what happened, with a screenshot path if captured.
  • Impact: Which visitors are affected, and on which task.
  • Remediation: A concrete change, naming the template or component to edit when it can be identified.
(repeat, ordered by severity; template-wide issues first)

Top Action Items (up to 5, by impact)

Top Action Items (up to 5, by impact)

  1. [SEVERITY] Description -- URL or template
  1. [SEVERITY] Description -- URL or template

Not Verified

Not Verified

Checks skipped for lack of a browser tool, access, or budget, so nobody reads silence as a pass.
Checks skipped for lack of a browser tool, access, or budget, so nobody reads silence as a pass.

Validate with Real Users

Validate with Real Users

A five-person, task-based test plan for the top tasks: the task prompts, what to observe, and what result would confirm or refute each HIGH or CRITICAL finding.
A five-person, task-based test plan for the top tasks: the task prompts, what to observe, and what result would confirm or refute each HIGH or CRITICAL finding.

Reference

Reference


Rank action items by how many visitors a fix helps on the top tasks, then by severity, then by effort. Keep recommendations proportionate to the site: a personal blog does not need an e-commerce checkout's rigor (the article's "When NOT to Focus on Usability").

某维度可以写为 `N/A — reason`。复审时,在 Score 后增加一个“Prev”列。

先按修复对主要任务中访客的帮助人数排列行动项,再按严重程度排列,最后按工作量排列。确保建议与网站规模相称:个人博客不需要达到电子商务结账流程的严格程度(参见文章中的“When NOT to Focus on Usability”)。