feature-demo-recording

Record a demo video of a web feature from a real browser. Two modes -- a NARRATED film where measured voiceover drives the timeline (designed slides, subtitles, punch-in camera, rendered from an HTML

SKILL.md

GitHub

Feature Demo Recording

Record a real browser and cut it into either of two things:

  • Narrated film (default when someone will watch and listen): designed brand slides, spoken narration, subtitles, and a camera that punches in on whatever is being talked about.
  • Silent evidence clip (a PR, a QA pass): the same capture with subtitles and punch-in, no audio, authored durations.

Both share one pipeline. The only difference is whether the timeline comes from measured speech or from durations you write down.

The rule that makes it work

Narration first, then record. Generate the audio, measure every line with ffprobe, and pace the capture to those numbers. Alignment is then true by construction instead of repaired in the edit. Recording first and narrating after accumulates drift you cannot fix without re-recording -- measured at about 3s by the last beat on a two-minute cut, versus under a few hundred ms when the recorder holds absolute targets read out of the measured timeline.

In silent mode there is nothing to measure, so you write each beat's duration in the script instead. Same file, same recorder, narrate.py --silent.

Pipeline

script.json ──narrate.py──> narration.mp3 + narr.json      (measured timeline)
              (--silent: no audio, durations taken from the script)
                                   │
                                   ▼  the recorder reads the beat targets
   record.py (adapted) ──> page.webm + events.json          (capture + beat log)
                                   │
                  ffmpeg ──> assets/footage.mp4
                                   │
      compose.py (narr.json + events.json) ──> index.html   (HTML timeline)
                                   │
                  hyperframes ──> renders/*.mp4
                                   │
      verify_align.py ──> drift / audio / picture / streams  (gate before delivery)

Only record.py is written per video. Everything else is generic.

How to invoke these scripts

Two rules come from the shared path-safety gate (references/_pathcheck.py), which every script under references/ routes its reads and writes through:

  1. Use an interpreter that can import kiro_crew. The gate consults the CENTRALIZED sensitive-path check rather than a local copy of the denylist, so when the package is not importable it fails closed with exit 78 rather than guessing. Any interpreter the package is installed into will do -- a provisioned worktree venv (<worktree>/.venv/bin/python) is the usual one but is NOT guaranteed to exist. Do NOT trust command -v kirocrew to name one: on a host using a version manager it resolves to a shim whose interpreter cannot import the package. Verify whichever you pick: <py> -c "import kiro_crew". Exit 78 with "cannot import kiro_crew" means the interpreter, not the script.
  2. Run from the video project directory, with paths inside it. Outputs must resolve within the working directory and inputs go through the descriptor-pinned read gate, so --out /tmp/x.html is refused by design.

The gate applies to references/ only. Your own recorder -- the copy of record_template.py that lives in the video project -- imports playwright and the standard library, and is not gated. It therefore needs the opposite thing: an interpreter with playwright installed, which is often NOT the one that can import kiro_crew. Two interpreters is the normal case, and a ModuleNotFoundError: playwright there is not the exit-78 gate.

PY=<worktree>/.venv/bin/python      # must satisfy: $PY -c "import kiro_crew"
REC=python3                         # must satisfy: $REC -c "import playwright"
cd <video-project-dir>
$PY <skill>/references/deps.py

What leaves the machine

Be straight with the user about this, because a demo often captures an internal dashboard. Narration goes through the same speech providers the product's own voice-reply path uses, so the answer depends on which one is picked:

  • The capture never leaves: Playwright runs headless and local, and the transcode is local ffmpeg.
  • --provider piper (preferred, what auto picks when it is installed): local neural TTS, so the narration text never leaves either. Needs the piper binary and an .onnx voice model.
  • --provider polly: synthesis happens in the caller's OWN AWS account through the aws CLI. Costs a little, needs credentials, stays inside their account.
  • --silent: no speech at all, nothing sent.

There is deliberately no third-party speech service. Narration text is product content and a demo often narrates an internal surface, so the only destinations this pipeline will speak to are the machine it runs on and the caller's own account. When neither is available it stops and says so rather than reaching for a convenient endpoint.

  • The renderer is fetched: hyperframes comes from npm on demand and the composition loads GSAP from a CDN, so a render needs network the first time.

references/deps.py prints which providers exist and which one auto will pick. For a strictly air-gapped run use piper or --silent, and pre-warm the npm/CDN caches -- do not claim the whole pipeline is local without checking.

Step 0 -- dependencies

$PY <skill>/references/deps.py            # report, exit 1 if anything is missing
$PY <skill>/references/deps.py --install  # install what is installable

Only user-level installs (pip --user, the Playwright browser cache, the npx cache). Anything needing root -- a language runtime, a system font -- is reported with the exact command for a human rather than escalated silently. Existing tooling on disk is reused, not rebuilt.

Dependency Why If missing
a speech provider narration (none needed in silent mode) reported, not auto-installed. piper counts as ready only with a voice model (KC_VIDEO_PIPER_MODEL), because resolve_provider needs one -- a binary alone would make the doctor claim a local default it will not actually pick. auto has no third-party fallback to reach for
ffmpeg with libx264 transcode the capture, normalise clips pip install --user imageio-ffmpeg (auto). The ffmpeg Playwright bundles is vp8/webm-only with no mp4 muxer, so the doctor checks for the encoder, not just the binary
ffprobe measures durations; the timeline depends on it NOT provided by imageio-ffmpeg -- needs a full ffmpeg install, so the doctor marks it human-installable
playwright + chromium drive and record the real UI pip install --user playwright + playwright install chromium (auto)
Node 18+ / npx runs the renderer reported, never installed -- this skill does not install a language runtime
hyperframes (pinned) HTML timeline -> MP4 fetched by npx at RENDER time. The doctor only asks the registry whether the pinned version resolves (npm view, metadata, nothing executed -- a check that runs a downloaded package is a worse hazard than the one it reports). It tries the configured registry FIRST and falls back to --registry=https://registry.npmjs.org when that one refuses (an authenticated mirror answers E401 for public packages), and tells you which applies
a CJK font only for CJK subtitles reported. An undeclared family fails the render gate rather than degrading, so the family name is configuration

Step 1 -- write and measure the narration

Copy <skill>/references/script.example.json, write the lines, then:

$PY <skill>/references/narrate.py script.json --out-dir assets/audio
$PY <skill>/references/narrate.py script.json --out-dir assets/audio --provider piper \
        --piper-model ~/.local/share/piper/en_US-amy-medium.onnx
$PY <skill>/references/narrate.py script.json --out-dir assets/audio --silent   # evidence clip

--provider auto prefers piper, then polly, and stops if neither is available -- there is no third-party fallback to fall through to. Check what that resolves to on the current host before running it on sensitive text.

Roles: intro (a designed slide), footage (over the capture), outro. say is spoken, cap is the subtitle -- keep them separate, because a glyph like the command key reads badly aloud and a subtitle must be shorter than a sentence. In silent mode give each line a dur.

The number of footage lines is the number of beats the recorder marks. That is the contract between steps 1 and 2.

Budget about 2.5 spoken words per second, and keep any single footage line under roughly 15s. A longer line leaves the frame sitting still while the voice keeps going, which reads as a dead shot; put long explanations on a slide, where a static frame is the intended look.

Step 2 -- record, paced by the timeline

Copy <skill>/references/record_template.py and adapt only the beat blocks. The template carries the parts that are easy to get wrong: absolute-target pacing, preroll_s capture, beat tagging, and cursor honesty.

hold_until(targets[n] - lead, page)   # idle until just before the beat
...do the work (click, type, navigate)...
hold_until(targets[n], page)          # absorb whatever the work cost
mark(page, "what is on screen", box, kind="focus", beat=n)

When the work cannot fit the lead -- a page reload on a short cut -- do the work and mark when the result is on screen, and say so in a comment. Watch the run log for ! overran beat target.

Selector note: production bundles often strip data-testid, so probe the running page for roles and accessible names instead of trusting selectors from source.

Step 3 -- transcode

FF is whatever deps.py reported for ffmpeg: an imageio-provided one is NOT on PATH, so a bare ffmpeg here would fail on a host the doctor called green.

FF=ffmpeg    # or the path deps.py printed for the ffmpeg row
"$FF" -y -i "$(cat MAIN_WEBM)" -c:v libx264 -preset medium -crf 20 \
      -pix_fmt yuv420p -an assets/footage.mp4

Step 4 -- compose and render

$PY <skill>/references/compose.py --brand brand.json   # -> index.html
npm run check     # in a hyperframes project; fix everything it reports
                  # a private mirror answers E401 for public packages; pin
                  # --registry=https://registry.npmjs.org in package.json
npm run render

Run check before every render. It has caught a silent-audio bug, overlapping tweens, an undeclared font and failing subtitle contrast in this pipeline -- each of which would otherwise have shipped.

Step 5 -- verify, then deliver

$PY <skill>/references/verify_align.py renders/<file>.mp4

Four checks, each for a failure that has actually shipped: per-beat drift against its line; audio loudness (a silent track reads about -91 dB; skipped in silent mode); picture luminance in slide and footage windows (a near-black footage window means the video element never seeked); and streams plus duration against the composition. Non-zero exit means do not deliver.

Then hand over the mp4 path, or file_send it if the user asked for Slack.

Scene-design rules

  1. Subtitles and slides are English-only unless the narration language is deliberately something else. They are our cards, not app content.
  2. Never let real user data into the frame. Record from an isolated pod, or create fresh demo data. A real session list carries chat titles and internal references.
  3. Pre-seed any state a beat depends on. If a beat needs a precondition, set it up in the pre-roll, which is outside the narrated timeline.
  4. Capture in dark mode when the brand frames are dark. Cutting from a designed dark slide to a white dashboard is jarring, and near-white subtitles over a light UI fail contrast no matter how heavy the shadow (the subtitle uses a solid pill, so readability holds either way).

House style, and where it comes from

This is worth stating plainly, because "brand slides plus a punch-in camera over screen capture" is a whole genre and a reader could reasonably ask whose look this is.

The palette is not a choice made here. useTheme.tsx sets DEFAULT_COLOR_THEME = 'kiro', so the Kiro theme is what a viewer already sees, and the brand frames use its tokens verbatim from the [data-theme="kiro-dark"] block in website/src/styles/theme-palettes.css: --bg:#19161d, --bg-elevated:#211d25, --accent:#8e48ff, --accent-hover:#9f63ff, --text-strong:#f2f1f4, --muted:#938f9b, --border:#352f3d. The capture is recorded on the same theme -- the recorder seeds mc-color-theme=kiro and mc-theme=dark -- so a slide and the dashboard behind it are one product rather than two. Getting this wrong is visible: an earlier cut here put emerald brand frames (the plain dark theme) around a Kiro-purple dashboard, and that mismatch is exactly what makes a film look assembled from parts.

The method is the distinctive part, not the look. Everything here follows from one rule -- the voice is measured first and the picture is paced to it. The genre norm is the opposite: record at a human pace, then compress the dead air afterwards. That inversion is what produces the properties below, and none of them are stylistic preferences; each is a number this pipeline paid for once.

  • A beat lands on its line, not near it. The recorder holds absolute targets read out of the measured timeline, so reload and typing overhead is absorbed. Budget: drift under a few hundred ms across a whole film. Duration-driven pacing measured about 3s of accumulated drift by the last beat, which is where this rule came from.
  • Roughly 2.5 spoken words per second, so a footage line over about 15s is a design error rather than a long sentence: the frame sits still while the voice keeps going. Long explanations belong on a slide, where a static frame is the intent.
  • A hold longer than 9s drifts instead of freezing, and the drift is sequenced strictly between the punch-out and the next punch-in, because two tweens on one property leave the value undefined on the overlapping frames.
  • A zoom that computes to 1.0 is a dead beat. Clamp to a minimum or frame a sub-region; a large focal element otherwise yields no push-in at all.
  • The cursor appears only for real pointer actions. A dot teleported onto a keyboard beat claims a click nobody made.
  • Subtitles sit on a solid pill rather than relying on a text shadow, so readability does not depend on what the captured UI is doing underneath.
  • Delivery is gated, not eyeballed. verify_align.py fails on per-beat drift, silence, a near-black footage window, or a duration that disagrees with the composition -- the same posture as a build gate, applied to a video.

What is genuinely inherited. The idea of punching in on the point that was just clicked is not original to this skill; it is common to the genre, and an earlier revision of this bundle carried an implementation derived from an external project. That implementation is gone -- easing now comes from GSAP inside the composition, and every constant above was set here. What remains shared with the genre is the concept, which is the level at which everyone shares it.

Landmines (each paid for once)

  1. An <audio> element with no id is not discovered by the renderer -- the file ships completely silent while every other check passes.
  2. events.json timestamps are on the footage clock, not the video's. The pre-roll happens before the beat clock starts, so the clip is offset by preroll_s. Getting this wrong puts voice and picture seconds apart with nothing else looking broken.
  3. Never let two tweens animate the same property at once. The long-hold drift is sequenced strictly between the punch-out and the next punch-in, or the scale is undefined on the overlapping frames and a seeking renderer bakes it in.
  4. A zoom that computes to 1.0 is a dead beat. A large focal element yields no push-in; clamp to a minimum or frame a sub-region.
  5. Only show the cursor for real pointer actions. A dot teleported onto a keyboard beat claims a click nobody made.
  6. npm's default registry may be an authenticated mirror -- pass the public one.
  7. Do not mint a dashboard credential from the recorder. A local safety policy blocks command lines pairing a product name with the word token, and moving that same mint into a child process routes around the control instead of satisfying it. The operator hands the finished URL in KC_VIDEO_TARGET_URL; the recorder only falls back to the bare pod URL, which is enough when the target asks for no credential.
  8. A FRESH pod shows first-run modals, and the gate is server-state driven: set the dashboard onboarded config as well as the localStorage keys. Changing pod config reloads the pod, so any credential in a URL you were handed is stale afterwards -- get the URL after the config change, not before. An already-provisioned pod needs none of this.
  9. The composition is generated, so fix the generator. compose.py writes index.html; a palette or layout change made in the generated file alone is silently reverted by the next compose, and -- worse -- a stale generator paired with a good index.html looks fine until someone re-runs the documented step. After changing either, re-compose and check the OUTPUT carries the change.
  10. The verifier pairs beats exactly only when the recorder tags them. Each primary mark carries beat=<index>; without the tag verify_align.py falls back to nearest-in-order inference, which reports a drift figure that is a guess. A recorder predating the tag pairs by inference and says so in its output -- believe that line before believing the number.

Not this skill

  • A still screenshot -> web-verify.
  • A quick GIF of one interaction with no cutting -> browser-recording.

Files

  • references/_pathcheck.py -- the shared path-safety gate every script reads and writes through (centralized sensitive-path check, fail-closed)
  • references/deps.py -- detect + user-level install + honest failure
  • references/narrate.py -- STEP 1: TTS, measure, emit the timeline (--silent supported)
  • references/compose.py -- STEP 4: narr.json + events.json -> index.html
  • references/verify_align.py -- STEP 5: drift / audio / picture / streams gate
  • references/record_template.py -- STEP 2: copy and adapt per video
  • references/script.example.json -- script format
  • references/brand.example.json -- palette, fonts, outro text