answer-me-with-html
A two-day-old skill that already does what most visualization skills promise: the model writes 923 tokens, the CLI draws the rest, and every number in the README checks out.
This is the most verifiable skill GearScope has reviewed in the visual-answer lane: the token-saving benchmark reproduces exactly from committed data, the 305KB bundled CLI works from a bare clone with no npm install, and the 219-test suite, byte-reproducible bundle, and five-job CI are disciplines most repos adopt years in, not on day two. The costs are youth and language: version 0.4.6 arrived in two days with no tags or changelog, the SKILL.md contract is Chinese-primary, its routing description exceeds the 1,024-character spec cap, and the flow component silently tolerates malformed edges that its sibling components reject with diagnostics.
$npx skills add QingYunA/answer-me-with-html --agent hermes-agent
lands SKILL.md plus the bundled CLI, byte-identical, 316KB
$/plugin marketplace add QingYunA/answer-me-with-html then /plugin install answer-me-with-html@answer-me-with-html
$git clone --depth 1 https://github.com/QingYunA/answer-me-with-html.git && cp -R answer-me-with-html/skills/answer-me-with-html ~/.claude/skills/answer-me-with-html
Node.js 20+ required, nothing else
install if
- Developers and team leads whose agents regularly explain architectures, protocols, or trade-offs. The committed benchmark shows the same answers for about a seventh of the output tokens and a quarter of the waiting, with cost flat, and the pages are single files that open offline.
- Chinese- and Japanese-language teams. First-class zh handling (wordlists, button labels, TTS voices) plus per-sentence language rules is rare in this lane, and the author's docs are fully bilingual.
- Anyone building explainer content with agents. The video mode (narrated, animated, MP4 export through locally installed Chrome and ffmpeg, no paid key required) is a genuine differentiator over page-only competitors.
skip if
- Users who want plain-text answers or audit every injected instruction in English. The SKILL.md body and CLI help are Chinese-primary, and the always-on mode attaches a roughly 120-token reminder to every prompt when enabled.
- Teams that need pinned, diffable releases. No tags, no changelog, no npm package, and a version cadence (0.4.6 in two days) that assumes you re-install to update.
- Hosts or strict spec environments. The 1,213-character description exceeds the 1,024-character Agent Skills cap; nothing enforces it today, but a strict host would truncate the routing text.
What It Does
answer-me-with-html is a single agent skill plus a bundled command-line renderer: when an answer is complex, the model writes a short extended-Markdown draft and hands it to a 305KB zero-dependency Node CLI (am), which lays out a self-contained one-page HTML explainer in about 50ms. The CLI ships inside the skill directory, so the only prerequisite is Node.js 20 or newer and there is no npm install step. Nine component types (flow with dagre auto-layout, sequence, tree, timeline, limits, annot, kv, callout, status tables) cover the shapes a complex answer takes, an ASD-STE100-derived lint checks the draft's prose in English, Chinese, and Japanese, a patch mode edits one panel of an existing page in place, and an am video mode renders 3Blue1Brown-style animated, narrated player pages with optional MP4 export. It targets developers and power users of coding agents (Claude Code, Codex, Cursor, OpenCode, Hermes via the skills CLI) who ask questions whose answers are diagrams, comparisons, and architectures rather than paragraphs. At review time the repo was two days old, at 720 stars and 52 forks, version 0.4.6, MIT, with 33 commits from four contributors.
The Good
Every quantitative claim in the README reproduces exactly from committed data. The README's headline table (6,873 output tokens plain vs 923 with the skill, 46s vs 13s, $0.22 vs $0.26 per answer, 7.4x fewer tokens) is derived from bench/results/results.json, 18 committed rows across three topics (tcp, cache, git) at three reps each on Claude Sonnet 5.5. Recomputing the documented way, per-topic medians then their mean, yields 6,873 and 923 tokens, 46.4s and 12.93s, $0.2196 and $0.2624, ratio 7.45x: every displayed figure matches its source to the digit, all 18 runs rendered on the first attempt, and the reproduction script ships in bench/. GearScope has not seen a younger repo with this bench discipline.
The zero-install claim is real and fast. From a bare clone with no npm install, node skills/answer-me-with-html/scripts/am.mjs renders a full page in 63ms of wall time (README claims about 50ms of render work). The two runtime dependencies, marked and @dagrejs/dagre, are bundled into the single file. The output page is self-contained: our English test draft produced a 27,555-byte file with lang="en" auto-detected, zero external CSS, JS, or font references, status words (ok/no/warn) rendered as glyphs, and the source draft embedded at #am-source for the copy-source and patch features.
Release engineering that most repos grow into, present on day two. The suite is 219 tests (216 pass, 0 fail, 3 skipped as e2e-gated) across 18 files covering components, parsing, patching, lint, robustness, and video. npm run build regenerates the committed CLI bundle byte-identically (md5 b8fd9c59dac030b9169b5537d57a8067), and CI enforces exactly that with a git diff --exit-code gate, on a matrix of Node 20/22/24 on Linux plus macOS and Windows, with coverage thresholds (85/85/80), an end-to-end video job that asserts an h264 1920x1080 MP4 through ffprobe, and an install-smoke job that runs the real npx skills add and Claude plugin flows. Both shipped plugins (the main one and the optional always-on companion) pass claude plugin validate --strict, and version 0.4.6 agrees across package.json, both plugin manifests, and the CLI banner.
Failure paths are designed, not accidental. A malformed sequence line is rejected with β L3 [sequence] plus the correct syntax example; timeline likewise. style: strict refuses to render a draft that fails the writing check, while deliberate strikethrough-marked examples are exempt as documented. The update check runs in a detached background subprocess with a 5-second timeout, a 7-day throttle, sends nothing but a GET of the project's own package.json, and turns off via config, CI detection, or AM_NO_UPDATE_CHECK. The always-on hook's reminder is silent the moment always: false is set (verified both states). Pages embed a minute-resolution generation stamp in the footer, and a cross-minute diff showed that stamp as the single changing line, so content is byte-stable in every part that matters.
The install rails all deliver, and the cheapest one is the best one. npx skills add QingYunA/answer-me-with-html --agent hermes-agent in an isolated HOME landed SKILL.md and scripts/am.mjs byte-identical to the repository, at a 316KB footprint, excluding the 21MB clone's demo videos and screenshot riders that a Claude plugin install would copy. The video path works end to end on a stock Mac: the committed TCP example rendered a 57KB player page in under a second without voice, system-voice narration (macOS say fallback when no ElevenLabs key is set) embedded base64 audio in 6s, and a two-scene custom draft exported an ffprobe-verified h264 1920x1080 MP4 through Chrome and ffmpeg.
The Bad
The routing description exceeds the Agent Skills spec cap. The SKILL.md description measures 1,213 characters against the 1,024-character maximum the spec and several hosts enforce. Nothing breaks today: claude plugin validate --strict passes and the skills.sh rail lands the file byte-identical. But a host that enforces the cap would truncate the description mid-trigger-list, degrading exactly the proactive-routing behavior the skill depends on. The same defect class was the number-one issue on the comparable onetake launch; here it is unflagged in the tracker.
The flow component is permissive where its siblings are strict. Sequence and timeline reject malformed lines with a line number and a correct example, but flow's grammar treats any bare line as a node declaration, so realistic agent typos render silently instead of erroring: A => B: fat arrow and A - > B: typo arrow each produced a single node named after the typo, and A ->> B: broken produced nodes "A" and "> B" with the label "broken", all exit 0. The README's "fixes its own mistakes" claim holds for half the component set; for flow, a typo yields a plausible-looking wrong page.
The SKILL.md contract is Chinese-primary. The frontmatter description is English, but the 198-line body an agent reads and executes, and all CLI help text, are Chinese. The bilingual README (20.5KB English, 20.4KB Chinese) mitigates this for humans, and models execute Chinese contracts fine, but non-Chinese-reading users cannot audit what is injected into their sessions without translation, which narrows the audience in a lane (visual explainers) where the incumbents are English-primary.
Two days old, versioned like a sprint, released like a prototype. 0.4.6 arrived within 48 hours of repo creation, with zero git tags, zero GitHub releases, and no changelog, so there is nothing to pin, diff, or roll back to; the npm package is prepared (bin entries, files whitelist) but unpublished (registry 404), leaving skills.sh and plugin marketplaces as the only distribution. Demand is correspondingly nascent: 220 installs on the canonical skills.sh entry (against 720 stars), one derivative fork already indexed, and one real bug filed on day two (issue #18: flow diagrams render all-black when CSS theme variables fail). The always-on reminder's cost is also understated: it measures 523 characters, roughly 120 to 130 tokens by the usual 4-chars-per-token heuristic, against the README's "about 90 tokens".
Smoke Test Results
Hands-on on the review host (macOS 26, aarch64, Node 24.13.1, Claude Code 2.1.152, ffmpeg 8.0.1). Scripts and full logs at sandbox/skills/answer-me-with-html/.
Run A. Fresh sandbox, no deps preinstalled
$ git clone --depth 1 https://github.com/QingYunA/answer-me-with-html
β
HEAD 300b530, pushed the morning of review, clone 21MB
$ node skills/answer-me-with-html/scripts/am.mjs help
β
banner "Answer me with HTML 0.4.6", full usage, no npm install needed
$ node .../am.mjs render - <<'EOF' # English draft, flow + status table
β
page written in 63ms wall, 27,555 bytes, lang="en", no external assets, #am-source embedded
$ node .../am.mjs render (same draft, second call)
β
byte-identical output (md5 match; the footer colophon stamp is the only cross-minute variable)
$ node .../am.mjs render # Chinese draft, then kana draft
β
lang="zh-CN" and lang="ja" auto-detected
$ node .../am.mjs render - # malformed sequence line
β
rejected exit 1: "β L3 [sequence] ... ζΆζ―εδ½ A -> B: ζ ηΎ" (line number, component, correct example)
$ node .../am.mjs render - # long sentence + "utilize" + passive voice
β
"STE 3 ζ‘θ¦ε" by default; style: strict refuses to render; strikethrough exempt as documented
$ node .../am.mjs patch page.html --panel "B Feature comparison" < new-panel.md
β
named panel replaced in place, other panels untouched
$ node .../am.mjs clean --dry-run && node plugins/answer-me-with-html-always/hooks/remind.mjs
β
clean previews deletions; hook emits the ~90-token reminder JSON when on, silent when always=false
NOTE (flow permissive grammar): a separate probe, "A => B: fat arrow", was accepted exit 0 as one node named "A => B", with no diagnostic
Pass rate: 9 of 9. The full harness runs 31 assertions (component list, config persistence, determinism, glyphs, lint detail, clean): 31 of 31 in the log.
Run B. Sandbox with deps preinstalled
$ npm install
β
clean in 1s warm (5s cold), 5 top-level packages, 14MB node_modules
$ npm test
β
219 tests: 216 pass, 0 fail, 3 skipped (e2e-gated video tests need AM_E2E=1)
$ npm run build
β
regenerates the committed CLI bundle byte-identically (the same gate CI enforces)
NOTE (npm rail): `npm view answer-me-with-html` returns 404; the package is prepared but unpublished, and the README does not claim npm
$ claude plugin validate /tmp/amwh --strict
β
passes for the main plugin
$ claude plugin validate plugins/answer-me-with-html-always --strict
β
passes for the always-on companion
$ grep version package.json .claude-plugin/plugin.json && node .../am.mjs help | head -1
β
0.4.6 consistent across package.json, plugin.json, marketplace.json, CLI banner
$ grep -c "git diff --exit-code" .github/workflows/ci.yml
β
CI gates bundle freshness, Node 20/22/24 + macOS + Windows matrix, coverage, e2e video, install smoke
Pass rate: 7 of 7. The npm 404 is recorded as an observation, not a failure, because npm is not a documented install path. CI on GitHub is green at HEAD (run #53, morning of review).
Run C. Functional verification (does it do what it claims?)
$ python3 amwh-bench.py bench/results/results.json
β
mean-of-per-topic-medians: 6,873/923 tokens, 46.4/12.9s, $0.2196/$0.2624, ratio 7.45x (README: 7.4x)
$ npx skills add QingYunA/answer-me-with-html --agent hermes-agent -y # isolated HOME
β
lands SKILL.md + scripts/am.mjs byte-identical to the repo, 316KB footprint
$ node .../am.mjs video examples/video-tcp.en.md --voice off --no-open
β
57KB player page in under a second; second render byte-identical
$ node .../am.mjs video examples/video-tcp.en.md --voice system --no-open
β
system-voice narration (macOS say) embedded as base64 audio in 6s, no API key
$ node .../am.mjs video custom-two-scene.md --voice off --mp4
β
h264 1920x1080 MP4 via Chrome + ffmpeg, verified with ffprobe (CI asserts the same)
NOTE (tracker): 4 open items on GitHub (1 bug, flow renders all-black when CSS theme variables fail; 1 image-support request; 2 PRs adding Chinese STE rules and a patch-style fix)
Functional pass rate: 5 of 5 prompts shown, 12 of 12 assertions in the harness log. The benchmark numbers are re-derived from the committed JSON, not re-measured with a live model; the render, lint, patch, video, TTS, MP4, hook, and install surfaces were all executed and asserted directly.
What the runs tell you
The install story and the renderer are production-grade today: everything works from a bare clone, the suite and CI are real, and both negative and positive paths behave as documented. The two FAIL-shaped observations that survived triage are a spec-compat risk (description length) and a parser-design choice (flow permissiveness), not breakage; two earlier FAILs during triage were harness bugs on our side (a too-strict lang marker grep, and pooling bench medians the README computed per-topic).
Setup Walkthrough
- Install the skill:
npx skills add QingYunA/answer-me-with-html --agent hermes-agent(or your agent's name; the installer supports more than 70). For Claude Code,/plugin marketplace add QingYunA/answer-me-with-htmlthen/plugin install answer-me-with-html@answer-me-with-html. - Ensure Node.js 20 or newer is on PATH. There is no other dependency; the CLI and both libraries it needs are bundled in the skill.
- Ask your agent a hard question ("explain the TCP three-way handshake", "Redis or Memcached"). The agent decides whether a page is worth it, writes the draft, and calls the CLI once; the page lands in
~/.answer-me-with-html/pages/and opens in the browser unless you turn that off. - Optional: install the always-on companion plugin if you want a small page attached to every conclusion, and say "config open off" if browser pop-ups interrupt you.
am clean --dry-runpreviews cleanup; nothing is deleted without consent.
Post-install gotchas: the Claude plugin route copies the whole repository including demo assets, so prefer the skills rail where footprint matters; and the weekly update check phones home to raw.githubusercontent.com for one version string (documented, no identifiers, easily disabled).
Alternatives
- nicobailon/visual-explainer (9.5K stars, GearScope 4.5/5) - turns tables, diffs, and plans into styled HTML pages after the fact; prefer it when you already have content to visualize rather than questions to answer; it carries npm dependency weight where this skill is zero-install.
- tt-a1i/archify (76.9K stars, GearScope 4.5/5) - validated diagram-first HTML with typed JSON IR and geometry gates; prefer it for engineering diagrams that must pass structural validation; it has no writing lint and no video lane.
- Asking the model for HTML directly - the plain mode in the committed benchmark: about the same cost per answer, 7.4x the output tokens and 3.6x the wait; prefer it only for one-off pages where installing anything is unacceptable.
Reviews stay honest because nobody pays us to publish them. If this one saved you time, throw a coin.
Tip the reviewer- reviewed by
- GearScope
- tested
- 2026-10-04 · macOS (Apple Silicon)
- last verified
- 2026-10-04
- depth
- HANDS-ON
- sponsorship
- none, ever
Don't install your next skill blind.
Every week: the shortlist of skills worth installing β and the ones to skip β from 100+ hands-on tests. No spam, no affiliate links.