Answer me with HTML: a page you can read instead of a wall of text

Ask your agent "Redis or Memcached for our cache?" with this skill installed and you get a comparison table with ticks, crosses and a verdict, on a page that opens in your browser. The model writes only a short Markdown draft. A CLI bundled with the skill handles layout, colors, dark mode and every diagram coordinate. It's released under the MIT license by its contributors, from GitHub user QingYunA's repository. Anthropic doesn't publish it. New to skills? Start with what Claude skills are.

Why the draft is short

Most of a hand-written HTML answer isn't content. Across 9 pages the author measured, a model writing HTML directly produced 4,893 output tokens on average against 612 with the skill, and took 31 seconds against 12, on Claude Sonnet 5.5. Nearly half the hand-written tokens were SVG coordinates. The bill falls by less, about 15% in that test, because every turn still reads the system prompt and the conversation.

What comes back

• How something works: flow charts, sequence diagrams and state machines, laid out by code so labels aren't cut off.
• A comparison: a table with ✓ and ✗ per row and a verdict at the end. Tables can be made sortable and filterable.
• A question about your repo: a folder tree and a call graph. Code blocks quote real lines from your files, and the CLI refuses files outside the folder and files that hold keys.
• A decision to make: option panels you pick on the page. The Reply button turns your choices and comments into one message to paste back.
• A one-line question: a one-line answer. No page.

Each page is a single .html file with no CDN links or web fonts, saved in ~/.answer-me-with-html/pages/, in one of three themes: blueprint, shadcn or paper. Drafts are checked against rules adapted from ASD-STE100, the controlled English used in aircraft manuals, for long sentences and passive voice. Ask for a video and am video builds a narrated player page instead; an MP4 export needs ffmpeg.

How to install it

You need Node.js 20 or newer, and there's no npm install step. Inside Claude Code:

/plugin marketplace add QingYunA/answer-me-with-html
/plugin install answer-me-with-html@answer-me-with-html

For Codex, Cursor or OpenCode, run npx skills add QingYunA/answer-me-with-html and pick the agents. Use one route per agent: the project's install guide warns that a skill and a plugin with the same name load twice and their instructions conflict.

Example request

Compare [option A] and [option B] for [our use, e.g. caching session data] on [cost, latency, setup effort, failure modes]. Answer as a page, [blueprint | paper] theme, and end with a verdict for [our situation].

To change behavior, run /answer-me-with-html:config, or just say "stop opening the browser". An optional always-on mode adds a rule to your CLAUDE.md or AGENTS.md so every conclusion gets a small page of 2 to 4 panels; it's off until you add that rule.

Limits

• It makes pages unprompted. The SKILL.md description says to use it "proactively and liberally", so expect pages you didn't request.
• Pages pile up. They stay in ~/.answer-me-with-html/ until you run am clean or ask the agent to clean up, and it asks before deleting.
• It checks for updates. The CLI looks for a new version on GitHub at most once a day and prompts at most every three days. The README says nothing about you is sent, and the update_check setting turns it off.
• No made-up numbers. Without real data it draws no charts, and any figure used only to illustrate gets labelled as illustrative.

When to use it

It suits anyone who asks an agent about architecture, trade-offs or a bug's causes and loses the answer in terminal scrollback. For a working app in one HTML file, use Web Artifacts Builder; for a standing map of a codebase, Floorplan.