Html Export

Codex compatibility note:

• Invoke $skill-name: load the registered .agents/skills/<name>/SKILL.md and execute it with Codex tools; no separate Skill tool or Claude mode switch is needed. Author .claude/**; mirrors are generated.
• Track steps before execution and keep progress/evidence current. Explicit skill requests execute their protocol. Use the native question tool for required questions.
• Map foreign tool names to available host capabilities. A name mismatch needs no approval; an actually unavailable required capability stops with its step/evidence and a user question before adaptation.
• Subagent authorization: when a skill is user-invoked or AI-detected and its protocol requires subagents, that skill activation authorizes use of the required spawn_agent subagent(s) for that task.
• Preserve protocol order unless the user approves a deviation. Workflows use $start-workflow: gates stay fixed; other steps may flex with logged reasons and evidence.

Quick Summary

Goal: Turn a local HTML file into PNG screenshots with a page-error and blank-capture check, a vector PDF, an MP4 or a GIF, through one entry point that reports the result as an exit code.

Workflow:

1. Pick a target — --to=png (render evidence), --to=pdf (print or slides), --to=mp4 / --to=gif (deterministic animation recording)
2. Check setup — dependencies are opt-in and skill-local; a missing one exits 3 with the setup commands
3. Export — node .claude/skills/html-export/scripts/export.cjs --to=<target> <input...> [options]
4. Read the exit code first — then report.json / frames.json, then the output; a written file is never proof on its own. A reused --out keeps earlier outputs: only the files this run's manifest names belong to it — report.json files[] (png), report.json output (pdf), frames.json output (mp4/gif)

Key Rules:

• --to is required; there is no default target.
• Nothing is ever installed on the user's behalf. Exit 3 means NOT VERIFIABLE plus a setup pointer.
• Render only HTML the project or the user produced; anything fetched or untrusted only with --offline (see Trust).
• The HTML stays canonical. Fix the page or pick the right selector; never restructure HTML for the exporter.

Be skeptical. Apply critical thinking, sequential thinking. Every claim needs traced proof, confidence percentages (Idea should be more than 80%).

Contents

• Quick Summary
• html-export
• Setup (opt-in only)
• Trust
• Usage
• Producer selectors
• Exit codes
• Calling from other skills
• What exit 0 proves
• Resources
• Closing Reminders

html-export

Agents and users may invoke this skill directly, and other skills may run the script by path. Either way the exporter only renders; it never installs, commits or publishes anything.

Setup (opt-in only)

Needs Node.js 20 or newer and, in the skill's own node_modules, Playwright 1.55.1 or newer (never a parent project's copy). Run these from the project root, only when the user asks; the exporter prints them on exit 3 and never runs them.

# Playwright and pdf-lib go into the skill folder; Chromium goes into Playwright's per-user browser cache
macOS/Linux shell:  (cd .claude/skills/html-export && npm install && npx playwright install chromium)
Linux system libs:  (cd .claude/skills/html-export && npx playwright install --with-deps chromium)
Windows PowerShell: Push-Location .claude/skills/html-export; npm install; npx playwright install chromium; Pop-Location
Windows cmd:        pushd .claude\skills\html-export && npm install && npx playwright install chromium && popd

# ffmpeg, only for mp4/gif (the build must include the libx264 encoder)
Windows:            winget install Gyan.FFmpeg
macOS:              brew install ffmpeg
Linux (Debian/Ubuntu): sudo apt-get install ffmpeg

Install from inside the skill folder, never with npm install --prefix: run from a project with its own package.json, that form makes npm add the host project as a dependency of the skill (it rewrites the skill's package.json and links the project root into its node_modules).

ffmpeg and ffprobe are looked up only in the absolute directories on PATH, never in the current directory.

Removing the skill. Deleting the skill folder removes Playwright, pdf-lib and every other npm dependency. Chromium stays in Playwright's per-user browser cache (or under PLAYWRIGHT_BROWSERS_PATH when that is set); to remove it too, first run npx playwright uninstall from inside the skill folder — it removes only the browsers no other Playwright installation still uses — then delete the folder. Regenerate the mirrors afterwards.

Trust

Render only HTML that the project or the user produced: its own decks, mockups, reports and drafts. For anything fetched from the web or otherwise untrusted, always pass --offline. Chromium runs without its OS sandbox, and a page can embed any local file the user can read (for example through an iframe) into the PNG, PDF or video. --offline blocks the page's network traffic; it is not a sandbox, and an embedded local file still lands in the output.

A page the project produced stays trusted when it uses fetched files only as verified images through <img src> (an SVG loaded as an image runs no script and loads nothing) and loads no remote script, frame or media; remote web fonts through CSS are fine. Inlining fetched SVG markup, or embedding it through <object>, <iframe> or <embed>, makes the page untrusted: pass --offline.

Usage

node .claude/skills/html-export/scripts/export.cjs --to=png page.html
node .claude/skills/html-export/scripts/export.cjs --to=pdf deck.html --page=1920x1080
node .claude/skills/html-export/scripts/export.cjs --to=mp4 intro.html --duration=4
node .claude/skills/html-export/scripts/export.cjs --help

A value flag takes =value or the next argument; --slides takes only the attached form, so --slides <css> reads the selector as an input file. Quote a selector with brackets or spaces in every shell: single quotes in bash, zsh and PowerShell (--slides='[data-state]'; unquoted, zsh treats the brackets as a glob), double quotes in cmd.exe.

The dispatcher checks every flag and input before any dependency or target runs, so a bad command exits 2 even when nothing is installed: an input, or a pdf --order entry, that is not an .html or .htm file (any letter case; a positional pdf input may also be a folder), an unknown flag or one the target rejects, --timeout or --scale not above 0, --timeout above 2147483647 (about 24.8 days), a malformed WxH, a size listed twice in --viewport, or an empty --slides=. Three checks need the page, so they run after rendering starts and stop the run at once: a --slides selector the browser cannot parse (png, pdf), a pdf selector that matches nothing, and a video with no duration. The video --keep-frames folder check runs after the dependency checks, before rendering.

Generic options

Browser start and stop. Launching Chromium, creating its context and page, setting up the --offline route, and closing the browser each get max(--timeout, 30000) ms, the same budget as the dependency check, so a short --timeout never fails a slow launch. A start call that runs over fails that input as a tool fault (pageFault: false, message <call> did not finish within N ms; the browser is not responding (a busy or broken machine, not the page)). A close that runs over prints Warning: the browser did not close within N ms; ..., the run continues with its exit code unchanged, and the exporter exits on its own instead of hanging; that browser is stopped when it exits.

--to=png — render evidence

Target flag: --full-page (opt-in). Inputs: one or more .html / .htm files; --self-check may run with no input to check only the verifier.

• Default viewports 1440x900,390x844. Each capture is viewport-only; --full-page adds <name>@<WxH>-full.png with the whole scrollable document and its own blank check. --full-page with --slides exits 2.
• Writes <name>@<WxH>.png, or with --slides one <name>@<WxH>-s<NN>.png per item, plus report.json.
• Log and continue: a failed open, navigation, capture or timeout on one input, viewport or slide goes to report.json failures[] (each entry with pageFault) and stderr, and the rest of the batch still runs. A page that fails to open still lists the errors it raised before that in errors[].
• Exit 4 on a page-caused failure, a blank capture, or a page error (unless --allow-errors), even when nothing was captured; a --slides selector that matches nothing counts as blank. Exit 1 only for a tool fault with none of those (browser start, a write error) or when --self-check reports verifier broken. A --slides selector the browser cannot parse exits 2 (report.json usageError). --page is a usage error; use --viewport.

--to=pdf — vector PDF

Target flag: --order=<file> — one .html / .htm path per line (blank and # lines skipped, relative paths resolved against the file's folder). Inputs: HTML files or directories (natural order), or --order, never both.

• Print mode (default): the page's print CSS. Page size: --page=WxH wins over a CSS @page { size } rule; without --page the CSS size applies, else Chromium's default.
• Slides mode (--slides[=<css>]): running animations are finished, then each navigated item prints as one viewport-sized page with screen media. --page with --slides is a usage error; size pages with one --viewport=WxH.
• Several inputs, or slides mode, merge into <out>/<name>.pdf with pdf-lib. stdout carries only the PDF path; report.json (errors, failures[], navigation, blocked[]) is written beside it.
• Log and continue: a failed input or slide is recorded (with pageFault) and skipped, and the PDF holds every page that printed. Exit 4 when the page caused a failure or reported errors (unless --allow-errors), even if nothing printed (then no PDF is written). Exit 1 when only the tool failed, or nothing printed without a page fault; the pages that printed are still written. A merge or PDF-write failure exits 1 with report.json fatal. A slides selector that cannot be parsed or matches nothing exits 2 with no PDF (report.json usageError). --self-check is rejected.

--to=mp4 / --to=gif — deterministic recording

Target flags: --fps=<1-60> (default 30 for mp4, 15 for gif), --duration=<seconds>, --audio=<local file> (mp4 only, AAC, trimmed to the shorter stream), --gif-width=<px> (gif only, default 960), --keep-frames. Input: exactly one HTML file.

• Units differ: --duration is in seconds; the page hook window.__duration is in milliseconds. --duration wins; with neither, the run exits 2 and writes nothing.
• Default viewport 1920x1080, one viewport only. --slides, --page and --self-check are usage errors. MP4 is libx264 at -crf 18; a build without libx264 exits 3.
• The recorder sets window.__recording before page scripts run and hides elements marked data-export-hide. Contract: references/animation-recording.md.
• Writes the video plus frames.json (status ok / page-errors / failed, exitCode, failure with pageFault, pageErrors, noMotion, a sha256 per frame). Page errors, including a timer or animation-frame callback that throws while the recorder advances the clock, are recorded and recording continues: exit 4 with the files written (0 with --allow-errors). A page that never loads or becomes ready, a __seek that throws, a page that breaks the recorder's setup or animation pinning, or any in-page call or frame step that exceeds --timeout exits 4 with no video. A tool fault (browser start, a write error, ffmpeg) exits 1, or 4 when the page also reported errors that --allow-errors does not accept. Identical frames print a warning and set noMotion (exit stays 0).
• --keep-frames also keeps every frame in <out>/frames/. An existing frames/ is reused only when it is a real directory holding nothing but the recorder's own frame files; otherwise the run exits 2 and touches nothing. Frames are captured into a work folder and replace the earlier frame files only after the video encoded, so a failed recording or encode leaves an earlier frames/ as it was. A failed run never writes the video path, and an earlier video stays in place; frames.json is rewritten by every run except a usage error.

Producer selectors

Any other page can mark its items with data-export-slide. feature-presentation slides carry both slide and deck__slide with a data-slide-id, so its existing section.deck__slide selector and the default selector find the same slides.

Exit codes

Calling from other skills

Invoke the skill, or run node .claude/skills/html-export/scripts/export.cjs ... by path. Every caller follows the same rule:

• Exit 0 → evidence, only within the scope below.
• Exit 4 → fix the page and re-run (fix loop); report.json or frames.json names each failure.
• Exit 3 → mark the check NOT VERIFIABLE with a one-line pointer to $html-export setup. Never run install commands on the user's behalf.
• Exit 1 / 2 → tool failure: quote stderr, mark NOT VERIFIABLE. Never count it as a design defect or a pass.
• Any other exit (such as 130 after an interrupt) → handle it like 1 / 2.
• Evidence files → only those the run's manifest names (report.json files[] for png, output for pdf, frames.json output for video); a reused --out still holds earlier outputs.

What exit 0 proves

Without --allow-errors:

• png: each input loaded at each requested viewport; no console error or uncaught exception from load through the last capture; no step failed; every capture held visible text, an image, an SVG or a canvas; with --slides, every matched item was shown and captured.
• pdf: every input and slide printed into the PDF and no page error occurred.
• mp4/gif: the video was encoded from frames captured on the recorder's controlled clock, with no page error.

It does not prove the page looks right, that interaction works, anything below the first viewport (unless --full-page), other browsers, or that remote assets load. Open the output and judge it; interactive checks stay NOT VERIFIABLE unless exercised another way.

Resources

• references/verification.md — png evidence, report.json fields, failure signatures, offline rule, self-check, known limits
• references/deck-export.md — pdf print vs slides mode, page size, per-producer commands, ordering, merge, pdf report.json
• references/animation-recording.md — the recording contract: time rule, window.__ready / __seek / __duration / __recording, data-export-hide, frames.json
• Tests: node .claude/skills/html-export/tests/run-tests.cjs (browser cases report ENVIRONMENT-BLOCKED when Chromium or ffmpeg is absent)

Closing Reminders

IMPORTANT MUST ATTENTION Goal: export HTML to PNG, PDF, MP4 or GIF through scripts/export.cjs --to={png|pdf|mp4|gif} and report the result by exit code.

IMPORTANT MUST ATTENTION read the exit code before the output — a written file proves nothing; exit 4 still writes files; only the run's manifest (report.json / frames.json) names this run's outputs IMPORTANT MUST ATTENTION never install dependencies on the user's behalf — exit 3 is NOT VERIFIABLE plus a setup pointer IMPORTANT MUST ATTENTION exit 1/2 is a tool failure, never a design defect or a pass IMPORTANT MUST ATTENTION --duration is seconds, window.__duration is milliseconds IMPORTANT MUST ATTENTION the HTML stays canonical — never restructure it for the exporter