Product promo, 1:1

The HTML video

dresser-teapot.html is the single-file export, playing right here. Click it, then press ? for the keys, or open it on its own.

Video: final.mp4 (final-6) · 1080x1080 · 30 fps · 20.00 s · 15.0 MB · -14.1 LUFS / -1.3 dBTP · Bumper: bumper.mp4 · 1080x1080 · 6.00 s · 2.6 MB · -14.1 LUFS / -1.5 dBTP · Exports: exports/final.square.mp4 (4.7 MB), exports/final.x.mp4 (1920x1080, 5.0 MB), exports/final.linkedin.mp4 (1920x1080, 5.0 MB) · Web page: dresser-teapot.html (one file, 6.7 MB, plays offline, no network requests)

The request#

"Make a 20-second square product promo from these museum photos of Christopher Dresser's teapot — sleek, but only facts from the museum record. Plus a 6-second bumper."

Mode: quick. The video will be published, so a researcher checked the facts and licenses before the build, and a critic reviewed the final. No questions were asked. The opening line stated the assumptions:

Quick mode: 20 s square master at 30 fps, plus a 6 s bumper cut down from it; no voice-over. On-screen words come only from The Met's record for object 823191. A lounge-jazz bed with its final hit on the end card, a thock on each spec row and a shimmer on the turn. Theme bold, with its ember accent swapped for silver. No Met logo. The end card points to the record, since the teapot is not for sale.

Contract: a promo-style film for a museum design object: Dresser's silverplate-and-ebony teapot, dated ca. 1879, shown in two views and a detail of its stamped marks. It uses only the facts The Met records (date, materials, size, weight, designer, manufactory) and ends on where to see the record.

What the research changed in the brief#

The researcher's fact sheet corrected the brief before any scene was built, and the film follows the fact sheet:

• Two views, not "three angles". The third photo (DP-18258-046) shows the stamped marks on the underside, not a view of the teapot. It became the opening image of the maker scene, captioned "Marks on the underside".
• "Ebony", not "Ebony handle". The record's medium is "Silverplate, ebony". The word "handle" is not in the record.
• Hook "ca. 1879.", not "Designed ca. 1879." The record gives the object's date. It does not give a design date.
• The record names a manufactory, James Dixon & Sons. The film names it but gives no founding year, because the record gives two different ones.
• bold is ember orange, not silver. --accent is overridden to #d4d2cf (13:1 on the ground).

The story, shot by shot#

Each scene starts on a downbeat of the composed bed (108 bpm, one bar = 2.222 s), rounded to the nearest frame, and each transition is centred on that downbeat (align center), so the visible change lands on the beat.

Bumper (6 s): the hook (0-1.5 s), the turn with its steam and shimmer (1.5-3.5 s), and a shorter end card (3.5-6 s: title, museum, credit, no-endorsement line). The bed is re-composed at 120 bpm so both cuts fall on beats, and both transitions are centred on them. The record URL is left out of the bumper: it could only be read for about 1.2 s there. The master's end card and share.txt carry it.

Commands, in order#

showtime is skills/showtime/bin/showtime. <job> is showtime-out/dresser-teapot-20260927-152814 and <p> is <job>/project. The Met's originals (4000 px, fetched by the researcher from images.metmuseum.org) are not shipped. The project has web-size copies.

showtime doctor --quick
showtime job init dresser-teapot --mode quick --goal "Make a 20-second square product promo ..." \
  --assumed "20 s master, 1:1 1080x1080, 30 fps; 6 s bumper cut down with retime -d 6" \
  --assumed "no voice-over: on-screen text from The Met record 823191 only; ..." \
  --assumed "two full views (025 side, 026 three-quarter) + underside marks detail (046)" \
  --assumed "theme bold with silver accent override; no Met logo; end card = record + CC0 credit + no-endorsement line"
showtime new dom <p> --title "Teapot, ca. 1879" --aspect 1:1 --duration 20   # native 1:1 layout; page then rewritten

# the object off its museum backdrop (macOS Vision here; rembg on Windows/Linux)
showtime assets cutout <p>/img/DP-18258-025.jpg --crop --pad 24   # "subject covers 17% of the image (vision, 3.2s)"
showtime assets cutout <p>/img/DP-18258-026.jpg --crop --pad 24   # 22 %, 2.1 s
#   edges checked at 100 % on the spout, handle end and feet; then resized to 1800/1600 px (teapot-025/026.png),
#   046 to 2400 px (marks-046.jpg), and two 256 px material swatches cut from 025 (silver body, ebony grip)
python3 <p>/tools/make_steam_lottie.py      # writes lottie/steam-curl.json (4 KB, hand-authored shape layers + trim paths)

# the bed: scene cuts must be downbeats
showtime audio compose --style lounge-jazz --dur 20 --sections "0:intro,3:verse,7:chorus,12:bridge,16.5:outro" --seed 1   # rejected, see Iterations
showtime audio compose --style lounge-jazz --bpm 108 --dur 20 \
  --sections "0:intro,2.2222:verse,6.6667:chorus,11.1111:bridge,15.5556:outro" --seed 1   # clean 1-2 bar sections, end_hit 17.78 s
#   -> the same spec in <p>/audio/mix.json, + shimmer (align peak) + 4 thocks (align hit)
showtime audio mix <p>/audio/mix.json -o <p>/work/mix.wav   # round 1: shimmer masked (-5.8 dB) -> duck + gain; intro +3 dB

# first look
showtime check <p>            # round 1: 1 contrast FAIL, 2 bottom-zone WARNs -> fixed; then PASS, 0 warnings
showtime snap <p> --every 1   # plus targeted --at sheets: 4 rounds (see Iterations)

# final, verify, review
showtime job note <job> --stage plan ... ; showtime job note <job> --stage first-look ...
showtime render <p> --job <job>          # final.mp4 (24 s for 600 frames); re-renders final-2 ... final-6
showtime qa <job>                        # final-6: PASS, 0 fail, 0 warn, 0 notes
showtime review-pack <job>               # round 1 (final-3) and round 2 (final-5), answered as self-review
showtime snap <job>/final-4.mp4 --at 13.33,14.2 --compare <job>/final-3.mp4

# the 6 s bumper: a copy of the project, three scenes, retimed
cp -R <p> <job>/bumper   # then: drop specs + maker, starts 0 / 2 / 4.6666666 (durations 2, 2.667, 3.333 = 8 s), bed at 120 bpm
showtime retime <job>/bumper -d 6        # hook 0-1.5, turn 1.5-3.5, end 3.5-6 (x0.75); data-at values scaled too
showtime check <job>/bumper              # PASS
showtime render <job>/bumper -o bumper-render/bumper.mp4
showtime qa bumper-render/bumper.mp4 --project <job>/bumper   # PASS

# post-ship polish (final-5 -> final-6, bumper), after the round-2 critic's "ship"
showtime check <p>                       # after centring the transitions: 1 contrast FAIL, 1 short-text WARN -> fixed; PASS
showtime render <p> --job <job>          # final-6.mp4 (30 s)
showtime render <job>/bumper -o bumper-render-r3/bumper.mp4
showtime snap <job>/final-6.mp4 --at 2.233,6.667,11.1,15.567,13.7,8.88,10.0 --compare <job>/final-5.mp4
showtime review-pack bumper-render-r3/bumper.mp4 --project <job>/bumper -o bumper-review   # the bumper's own pack

# deliver
showtime deliver exports <job> --targets square,x,linkedin
showtime qa <job>/exports/final-6.x.mp4 --platform x           # PASS (also linkedin, square: PASS)
showtime export html <p> --target artifact --subtitle "Christopher Dresser · The Met, object 823191" -o html/dresser-teapot.html
showtime clean <job> -y                  # freed 105.6 MB (bumper: 1.8 MB)

To render the bumper from this folder, copy project/img/ and project/lottie/ into bumper/ first. They are not duplicated in the repo.

Iterations that mattered#

1. The bed's grid set the cut list. At the style's own 112 bpm, with cuts at 3/7/12/16.5 s, the composer raised the tempo to 120 and inserted 2-beat bars to land the markers. At 108 bpm a bar is 2.222 s, so the scenes were set to whole bars instead: hook 1 bar, then 2 bars each for turn, specs, maker and end. The hook is shorter than the brief's 3 s (2.23 s), which fits the story rule that the first cut should come by about 2 s.
2. End card (check round 1). The end teapot rendered at its natural 1600 px behind the title, and check reported a contrast FAIL (1.21:1 at 16.23 s). It is now sized to 38cqw. "1.1 lb." and the credit line sat in the bottom 8 % of the frame (the player UI zone), so the rows and the credit moved up.
3. Sheen. The first version was a 24 %-wide band at 0.55 alpha that washed out the whole teapot at 3.2 s. It is now a 12 % band at 0.42.
4. One-frame flash at two cuts (final-2 → final-3). The review pack's cuts.jpg showed the bare incoming scene (only its label, the grid not yet in) for exactly one frame at 6.667 s and 15.567 s, and then the outgoing scene again. Cause: clips snap data-start to a frame when it is within 1 ms, but CSS transition windows test t >= start - 1 µs. So 6.6667 made the clip active at frame 200 while the slide had not started yet. The scene starts are now written just below the frame time (6.6666666) (see the project's head comment). Neither check nor qa flagged this.
5. End card motion. qa noted a 2.6 s final still hold. The end teapot now drifts slowly (scale 1 → 1.06), and its sheen peaks on the bed's final hit at 17.78 s.
6. Maker pill (critic polish, final-3 → final-5). "Marks on the underside" covered the top of the registration diamond as the push rose. The push now ends 9 % to the right, and the pill leaves at 13.65-14.0 s, before the diamond reaches it. In the post-ship polish (item 7) the fade moved to 13.5-13.85 s.
7. Post-ship polish (final-5 → final-6, bumper). A second critic pass found the visible transitions starting 0.11-0.23 s after the downbeats their scenes start on (the iris ring first showed at 2.40 s for a 2.233 s cut). All four transitions are now centred on their cut (data-transition="slide left 0.6 center"), so each is mid-way on the beat; the scene starts, SFX and bed did not move. Centring moved check's sample times, and it then caught two things the old samples had missed: the dimmed "Ebony" row's label was 3.74:1 at 10.0 s (the rows now dim to 72 %, not 55 %), and the manufactory line lost 0.3 s of readable time to the earlier iris (it now fades in at 12.9 s, not 13.2 s). The row labels read "Material", not "Medium", which a stranger could read as a size. The bumper drops the record URL (see above).

Review#

No sub-agent tool was available in this session, so the maker answered the critic brief itself. Round 1 (review/round-1-FINDINGS.md, on final-3) gave "ship after fixes": no blockers, no should-fixes, and one polish item (the pill overlap), which is fixed. Round 2 (review/round-2-FINDINGS.md, on final-5) checked the fix and gave "ship". Every on-screen word was checked against the researcher's allowed-wording list. After "ship", a second critic pass on final-5 (with the bumper seen through ffmpeg frames) found no blockers, one should-fix (the bumper's URL was readable for only ~1.2 s) and four polish notes: transitions lagging the downbeats, the pill's last opaque frames ~9 px from the diamond, "Medium" readable as a size, and a one-frame iris artefact in the bumper. The first three and the should-fix are done (iteration 7). The iris artefact is normal iris geometry and is left as is. Each fix was checked with before/after snaps at the cited times. The bumper then got its own review pack, answered as a self-review ("ship"). That pass also noted that review-pack's cuts.jpg samples around data-start (-2f..+4f) and so can miss transitions that start later; with the transitions now centred, the bumper's cuts.jpg shows them mid-way. A human look and a listen are still welcome, because this session can meter audio but cannot hear it.

Features shown#

assets cutout (macOS Vision, two cut-outs; rembg is the fallback on other OSes) · Lottie (ST.lottie seeking an original JSON: three stroked shape layers with trim paths, generated by tools/make_steam_lottie.py; frames at 4.2 s and 4.8 s are pixel-identical whatever the seek order) · feature-grid (1 column, image and Lucide icons, focus walk) · ken-burns (the underside push) · kinetic-type (slide) · end-card (mask, tagline, CTA, URL) · theme bold with a silver --accent · transitions sdf-iris (WebGL), slide ×2, iris, all align center · --aspect 1:1 native layout · audio compose lounge-jazz (sections on the cuts, section_gain, end_hit used for the end-card sheen) · audio sfx thock ×4 and shimmer (align: peak, the bed ducked under it) · retime -d 6 cut-down · expect.must_show ("ca. 1879", "500 g") · check, snap (project, video, --compare) · render --job, qa (+ --platform x|linkedin|square) · review-pack ×2 · deliver exports --targets square,x,linkedin · export html --target artifact · clean · researcher (claims and licenses), critic (self-review).

QA summary#

• final.mp4 (final-6): PASS (0 fail, 0 warn, 0 notes). h264 High yuv420p, 1080x1080 30 fps, faststart, BT.709, 20.00 s, -14.1 LUFS, -1.3 dBTP. No silent gaps, no black or frozen stretches, and frame 0 flows into frame 1. No attribution is required. must_show "ca. 1879" is on screen 0.00-1.87 s and "500 g" at 7.37-11.23 s (the centred transitions start earlier).
• bumper.mp4: PASS (0 fail, 0 warn), 6.00 s, -14.1 LUFS, -1.5 dBTP, must_show "ca. 1879" at 0.00-1.20 s.
• Exports: square (1080x1080, 4.7 MB), x and linkedin (1920x1080, the square picture pillarboxed on a blurred copy, 5.0 MB each). All three PASS against their platform. Both X and LinkedIn also accept the square master as it is.
• check: PASS, 0 errors, 0 warnings (master and bumper). The bumper has one info note: sub-pixel anti-aliasing at 0.9, 3.9 and 5.4 s, 0.00 % of pixels.
• Mix: shimmer +4.2 dB above the bed, thocks +6.2 to +11.7 dB, and the bed's final hit at 17.78 s.
• HTML: 6.7 MB of the 16 MB limit, embedded AAC 96k at -14.1 LUFS, 5 chapters (one per scene), and no network requests.

Sources and licenses#

Not claimed anywhere: "iconic", "modern", "ahead of its time" or any other adjective as fact; a design year; a decoded registration date; a founding year for the manufactory; "on view" (true on 2026-09-27, but it can change); price or availability. The courtesy credit is in credits.txt. Post copy is in share.txt.

The project files on GitHub