Build an animated, data-driven explainer film for a real system — a canvas film that renders live from the product's own data and cannot drift, not an exported video. Can additionally stamp the finished film into a voiced MP4 (deterministic frame render + timed ElevenLabs narration) for channels that only take a video file. Use when someone wants to explain how a system works visually, wants an explainer/hero video, references a motion-graphics explainer to imitate, or asks to show a mechanism rather than describe it. NOT for static diagrams or single illustrations.
Build an animated, data-driven explainer film for a real system — a canvas film that renders live from the product's own data and cannot drift, not an exported video. Can additionally stamp the finished film into a voiced MP4 (deterministic frame render + timed ElevenLabs narration) for channels that only take a video file. Use when someone wants to explain how a system works visually, wants an explainer/hero video, references a motion-graphics explainer to imitate, or asks to show a mechanism rather than describe it. NOT for static diagrams or single illustrations.
category
visual-communication
argument-hint
<system or topic to explain> [--reference <video-url>] [--data <path>] [--export] [--voice]
{"version":"1.26","created":"2026-07-15T00:00:00.000Z","author":"Ability.ai","changelog":["1.26: Promoted to the public trinity-skills library — declared requires:/argument-hint, documented cold-start behavior for the optional ElevenLabs keys (the film itself needs no credentials; only the Step 11 voiced export does), skill-relative script paths, sibling-skill references made optional, and build provenance genericized for a public repo","1.25: WSC caption-band trap (multi-agent concepts film): the caption-safe world-y rule (cy + (0.79-0.5)/z) ignores the world-scale inset (WSC 0.875), so bottom-row structures thought cropped at z 1.2-1.4 render straight under the captions in every mildly-zoomed shot; fix = include WSC in the safe-y math, or focus-dim the bottom row during chapters where it is furniture (reference.md → Layout traps)","1.24: Standing-fixture accumulation trap (user-reported on the BE intelligence-layer film): in films that BUILD UP a persistent world (stack/map films), each chapter's vignette furniture (annotation chips, helper persons, shelf labels, dashed marks) silently accumulates into overlap clutter by the late chapters — and every per-beat review passes because each chapter looks fine in ITS OWN frames. Rule: transient annotations fade when their chapter ends (keep only core structures); verify by reviewing LATE-chapter frames for accumulated fixtures and transition midpoints, not just each chapter's own beats. Also: if a closing frame draws a boundary (ownership/sovereignty metaphor), assert every fixture's extents sit INSIDE it — side panels poking through the frame edge break the metaphor","1.23: Export top-left/ghost-margin trap (user-reported twice, misdiagnosed once as composition): export-film.js shrinks the canvas CSS width and waits for a re-mount, but films whose engine uses addEventListener(resize, fit) never re-mount (inline style change fires no window resize) — seek() then paints the new smaller CSS size into the stale load-time backing store: content top-left at ~77%, stale pixels ghosting in the margins, while browser/check-page/shoot-beats all look perfect. Fix in exporter: dispatch new Event(resize) after the style change AND hard-fail if the store width != --width. The film-side symptom is identical to a composition bug — check meta.w BEFORE re-authoring layouts","1.22: Async-seek export trap from a 12-film course batch — an engine whose seek() only sets state (paint on next rAF) is invisibly broken for export-film.js, which batches 12 seek+toDataURL calls per evaluate() with no rAF between them: every batch captures 12 stale frames (~2.5fps output) while the browser, check-page, and shoot-beats all look perfect. seek() must draw synchronously (the exporter's contract); audit with exact-frame-hash uniqueness on an animating segment, not mpdecimate (which merges quantized dark-fade steps)","1.21: trailing_silence substring bug — 'silence_start: 0' in line also matches 0.523, so a MID-CLIP pause could be returned as the tail and the exact-duration cut truncated the line's end by that amount; parse numerically (start ≈ 0 only). Also learned: v3 clips end HOT (zero trailing silence), so most cuts = full clip length","1.20: Tail gate −45dB → −60dB — a final word's quiet decay counted as silence at −45, so the exact-duration cut (1.19) truncated it mid-word ('…Yo'). Only provable silence (true dead tail ≈ −90dB) may trim; longer effective durations are the fit machinery's problem, never the listener's","1.19: The render trim now IS the measurement — atrim at each clip's measured effective duration (declick fade at the cut) replaces silenceremove. Trap it fixes: silencedetect (absolute) and silenceremove (windowed RMS) disagree on v3 breathy tails by ~0.3s, so the render kept tail the fit math thought was cut — the min-gap guarantee silently shrank (measured: 0.15s of real air where 0.3s was enforced)","1.18: voice-film.py --base-speedup — uniform render-time atempo on every clip (the global pace lever for models that ignore the API speed knob, e.g. eleven_v3; 1.05 = 5% faster, zero re-billing); fit report and spill math now use effective (atempo-adjusted) durations","1.17: Guaranteed inter-line breathing gap — voice-film.py --min-gap (default 0.3s): each line must END that much silence before the NEXT line's start; line starts stay pinned to caption times (the film alignment), the previous line absorbs the gap via auto-atempo or a trim. Matters most on eleven_v3, whose dense expressive read can otherwise butt lines together with ~0.05s of air","1.16: Default TTS model → eleven_v3 (most expressive tier). Two measured traps: v3 accepts voice_settings.speed (HTTP 200) but ignores it — pace v3 with --max-speedup 1.1 + text trims (pin model_id eleven_multilingual_v2 when the speed knob is required) — and v3 rejects previous/next_text (400 unsupported_model), so voice-film.py now auto-skips prosody context on v3","1.15: Terminal-spike trap from a course film — ElevenLabs clips can end in a sharp click/chirp on the final word, and the tail-trim removes only silence so it survives the render; voice-film.py now fades the last 0.12s of every clip at its trim point (afade=t=in between the areverse pair — declicks the tail and softens the ducking release). Existing spiked clip: evict from vo_clips/ + re-roll; verify with 100ms max_volume slices (must ramp monotonically to silence)","1.14: From the knowledge-agent film re-cut — verbatim-subtitle mode is now the default for voiced films (one line, one font, byte-identical to narration.json + scripted parity check; user-confirmed that caption/voice divergence reads as an error); global pace as a single t/SHORTEN divisor (reference: Film skeleton); four new traps — scrim must outlive its scene's contents (previous-scene flash), camera-gated world labels, shoot-beats integer-second filename collisions, and true-silence-first ducking measurement with the -v error / -ss-after-i ffmpeg gotchas","1.13: Trap from the knowledge-agent lens re-export - inserting a narration line invalidates the index-keyed clip cache for every later line and re-rolls their durations; re-read the whole fit report after any insert","1.12: Container-offset trap from the builders-film export — voice-film.py MP4s carry audio start_time≈0.976 (edit list); WAV-flattened onset checks read ALL lines a uniform ~1s early. A constant bias = container offset, never a placement bug; verify on the muxed MP4 (PTS honored) with a music-floor-aware gate + windowed volumedetect","1.11: Trap from the knowledge-agent music export - the onset check is blind on a music mix; verify placement on a no-music mux of the same clips and prove ducking separately","1.10: Onset-check trap from the knowledge-agent export — a soft-spoken FIRST line can sit under the silencedetect gate and read as missing; prove placement with volumedetect (speech-level window vs −90dB gap), never by lowering the gate","1.9: Trap from the builders-film build — shoot-beats' scrub-click seeking drifts (pixel-quantized), producing phantom timing bugs; verify timing-sensitive frames via the __films seek hook + a getImageData probe before touching the draw code","1.8: Voice pacing + honest fit math in voice-film.py — `speed` (ElevenLabs voice_settings, 0.7-1.2) fixes a slow narrator globally BEFORE any words are cut (Daniel at 1.15 cleared 21 of 25 collisions untouched), and every clip's ~0.5s dead tail is now measured, excluded from the fit report, and trimmed identically in the render","1.7: Background music in the voiced export — voice-film.py --music FILE (loop/trim/fade + sidechain ducking under the voice). Trap: sidechaincompress silently does nothing at default gain staging — the sidechain needs level_sc boost to clear the threshold; verify ducking by measuring a speech span vs a gap, never by ear alone","1.6: Voiced video export — scripts/export-film.js (seek-hook frame render → MP4) + scripts/voice-film.py (timed ElevenLabs narration with a collision-checked fit report, muxed). Proven on the Trinity film (195s, 29 lines). The film stays the source of truth; the MP4 is a dated stamp","1.5: Two layout traps from the Trinity business-explainer build — data-driven labels need measureText fit-to-width (one font size can't fit real entity names), and stage→overview transitions need overlapping fade windows (screenshot transition midpoints, not just beat centres)","1.4: Trap from the Trinity extension: scripted time-shifts must keep sub-threshold literals byte-identical (no reformatting) and assert every later anchor before writing","1.3: Layout trap from the knowledge-agent orb-restyle — a background-attachment:fixed page gradient paints only the viewport band (white elsewhere in stitched full-page captures and on iOS Safari); fix is a solid base color on html under it","1.2: Review traps from the Trinity build — hairline strokes read bright in downscaled contact sheets (measure with getImageData, not eyes), and a zero-install Playwright-MCP rig (http.server + window.__films seek hook + in-page composite contact sheets)","1.1: Four render-only traps added to reference.md from the knowledge-agent build — draw-helper globalAlpha reset silently defeating label fades, missing charset meta producing canvas mojibake, camera-clamp shifting zoomed framings so world-space labels clip or hit the screen-space caption band, and radial arcs bunching/wrapping labelled items (use a list + spoke)","1.0: Initial version — distilled from building a forecasting platform's hero explainer end to end (reference decode → data interrogation → worked example → canvas film → React port → shipped live)"]}
Animated Explainer
ℹ️ First, set expectations: before anything else, print one short line with this skill's version and its most recent change — the top entry of metadata.changelog above — e.g. animated-explainer vX.Y — recent: <summary>. Then proceed.
ultrathink
Purpose
Build an animated explainer of how a real system works — rendered live to a <canvas>
from the system's own data, so it regenerates on every publish and can never drift from
what the product reports.
The thesis: a data-driven film, not a video. An exported MP4 is stale the day the data
moves, and it quietly becomes a lie. A film that reads data/*.json at render time changes
when the system changes. (This is testable: on the forecasting build, a data refresh landed
mid-deploy and the calibration curve visibly redrew.)
Use this when someone wants to show a mechanism — how a system works, why it's
trustworthy, what happens to one item as it moves through.
Don't use this for static diagrams or illustrations — a diagram skill (if your
workspace has one) is the right tool there. This skill is specifically time-based motion
driven by live data.
State Dependencies
Source
Location
Read
Write
The system's real data
product repo data/*.json (or equivalent)
✅
❌ never
The system's source of truth
ARCHITECTURE.md, papers, specs — not published artifacts
Final step, only if the film ships to a live site and the user has asked for a deploy. Optional — this skill never requires it.
Credentials & cold start
Building, verifying, and shipping a film needs no credentials at all. Only the optional
voiced export (Step 11) reads any:
Key
Needed for
Missing →
ELEVENLABS_API_KEY
Step 11 TTS narration
Fail fast with ELEVENLABS_API_KEY not set — add it to this agent's credentials; every earlier step still runs, and the silent MP4 export is unaffected.
ELEVENLABS_VOICE_ID
default voice for Step 11
Optional — voice_id in narration.json or --voice overrides it. Absent with no override → the tool stops and names the three ways to supply it.
PLAYWRIGHT_DIR
pointing the render scripts at an existing playwright install
Optional — the scripts resolve playwright from the caller's cwd first, then this variable, then the skill directory.
Script paths below are relative to this skill's directory; run them from there, or
prefix with the directory the skill was injected into.
Core principles
These are load-bearing. Each one was paid for.
1. The data is the story's boss — interrogate it before you design the arc
Decide what's true before you decide what's compelling. Concepts must die here; that
is the step working, not failing.
Two died on the forecasting build, both after the storyboard felt great:
"watch it get better over time" — the metric was flat (0.217 → 0.188 → 0.210). Noise,
not learning. Animating it would have faked the one number the brand rests on.
"performance improving" curve — the sample was back-loaded (Q1 spanned 3 months, Q4
spanned 1 day), so the apparent decline was a sampling artifact, not improvement.
Run the numbers yourself. Check the shape and the sampling, not just the headline.
Finding this at storyboard costs an hour; finding it after a render costs a week; not finding
it costs your credibility.
2. Read the source of truth, not the published artifacts
Published data dirs are a curated snapshot of a system, not the system. On that build the whole
first film was built from the website's data/ and got the architecture wrong — the real
mechanism was in ARCHITECTURE.md and an unpublished paper. Find the design docs.
3. One concrete worked example beats any amount of abstraction
Follow one real item end to end. The film only worked once it became a single real
forecast — said 88%, world said no — traced from question to consequence.
Prefer a failure. A miss shows the machine doing its job, is impossible to read as
cherry-picking, and gives every downstream loop something real to react to.
4. Explainer, not advertisement
Assume the viewer knows nothing. It must read with a voiceover on top and without one.
Open on the problem, not the architecture — state what's broken, then reveal the machine
as the answer. A hook is not a flourish; it's the reason to keep watching.
5. Separate the mechanism from the numbers
Statistics clutter an explanation. Carry the mechanism with symbolic marks, then put real
figures at the very end — or nowhere. If a mark is not 1:1 with a thing, say so on screen
(1 dot ≈ 30 beliefs). If something is too rare to be field scale (17 of 44,510 = 0.6 of a
dot), draw it 1:1 as an annotated callout and label it — never silently mix scales.
6. Nothing teleports
If a thing appears somewhere new, animate it getting there. A chart that materialises is a
cut; a chart that dots visibly fall onto is an explanation. This was the single biggest
"it doesn't explain it better" complaint on that build, and fixing it changed the film.
7. Lock the layout
One circuit. Nothing crosses anything. Spaghetti beziers read as complexity, not as a
system. A horizontal spine + a return rail + entities sitting on the rail reads instantly.
8. Show the part that doesn't work
That system's Loop 0 stays dashed and hollow for the whole film because it has commissioned
consequences but zero resolved. A diagram where everything works is the least trustworthy
thing you can draw. The unfinished part is the credibility.
9. Screenshot every beat. Measure anything computed. They are different jobs.
Look at renders → catches collisions, clipping, off-frame labels, composition.
Measure in the browser/numerically → catches wrong values, wrong layout, aliasing.
Neither substitutes for the other. On that build, every numeric bug was invisible to the eye and
every layout bug was invisible to the code.
10. Lift proven code — never retype it
Port by script, not by hand. Hand-copying a data payload once silently dropped a record
and changed a headline number from 0.194 to 0.195.
Process
Step 1 — Decode the reference (if there is one)
Never design against a description of a reference. Get the artifact.
yt-dlp -o ref.mp4 "<url>"# X/social links often block WebFetch
ffmpeg -i ref.mp4 -vf "fps=1/2,scale=520:520,tile=4x5" sheet.png # contact sheet of the arc
ffmpeg -ss 12 -i ref.mp4 -frames:v 1 f.png # a specific beat, full res
Then sample its actual palette (don't eyeball hex): see reference.md → Palette sampling.
Write down what the reference does, decision by decision, and explicitly mark which
decisions transfer and which you'll break — and why.
Step 2 — Interrogate the data (the gate)
Before any storyboard, answer in writing:
What does the data actually support? What does it refuse to support?
Is any apparent trend an artifact of sampling (uneven time buckets, back-loading)?
What's the most interesting true thing here?
If the story you wanted isn't supported: change the story, say so, and offer the honest
alternative. Do not soften it into an animation.
Step 3 — Find the source of truth
Read the architecture docs, papers, specs. Not the published JSON. Reconcile disagreements and
note which snapshot each figure comes from (dates matter — never mix snapshots in one frame).
Step 4 — Choose the worked example
One real item, ideally a real failure. Pull it verbatim: the claim, the number, the outcome,
the reasoning. Note precisely what is verbatim and what you compose — and disclose it.
Step 5 — Storyboard the beats
Problem → question → commitment → reality → "but does that prove anything?" → mechanism →
payoff. Write the captions first: if the captions alone tell the story, the film will work.
Decide the caption mode now, not at export time:
Voiced film planned (the usual case) → one caption line, one font, verbatim what the
voice will say. Viewers read while they listen; a second font or diverging text reads as
an error, not as extra detail (user-confirmed on the knowledge-agent film — "having two different
fonts is confusing; what the voice is saying should exactly match the subtitles"). Fine
detail (names, edge types, figures) lives in on-canvas diagram labels, not a caption sub-line.
Silent-only film → caption pairs (main + sub) are fine; the sub-line carries the detail.
Step 6 — Build in a standalone HTML
One self-contained file, no build step, opens in a browser. Fast iteration is everything here.
Use the product's real brand tokens. See reference.md for the film skeleton, camera,
path-following, and growth recipes.
Author the storyboard on its own internal clock and treat overall pace as a single divisor
(t = t / SHORTEN, first line of draw) — "make it 20% faster" then costs one constant, not
a re-authoring of every beat. Recipe: reference.md → Film skeleton.
Delete old screenshots before every shoot (rm -f shot_*.png) — stale frames polluted
three separate reviews and produced two phantom "bugs".
Look at every frame. Then measure anything numeric (see reference.md → Measuring).
Step 8 — Port to a component (by script)
Lift the draw block out of the proven HTML with a script. Hoist constants to module scope.
Feed data in as props from the server component, computed from the repo's own files — this
is what makes it self-updating. See reference.md → React port (and its effect-structure
trap, which silently breaks scrubbing).
Step 9 — Build-gate, then drive the real page
npm run build # never ship a broken build
npm run start # then drive the REAL page, not just the static HTML
Kill stale servers first (kill -9 $(lsof -ti:3000)) — testing a stale build wastes a cycle.
Assert: canvas mounts at expected size, no JS errors, no horizontal overflow, scrubbing
lands where you click.
Step 10 — Ship (only if asked)
Invoke /publish-site. Deploy source only; don't bundle a data refresh into a component
deploy — it muddies rollback. Verify the live page by driving it, not by curling for a 200.
Step 11 — Export a voiced video (only if asked)
The film is the source of truth; some channels (YouTube, LinkedIn, X) only take a video file.
This step stamps the film into a voiced MP4. The stamp is dated the moment the film or its
data changes — say so wherever it's posted, and re-export instead of editing the video.
# 1. Render frames deterministically through the film's own seek() hook → silent MP4
node scripts/export-film.js film.html --out film_silent.mp4 # 1080p30 by default# 2. Author narration.json — in verbatim mode (the default, Step 5) each line is the caption# text byte-for-byte; in compressed mode it's the caption spine, trimmed to fit.# Either way: never introduce claims that aren't on screen.# 3. Generate clips + fit report (no render). Iterate on the text until 0 collisions.
python3 scripts/voice-film.py narration.json --film film_silent.mp4 --report-only
# 4. Render: timed voiceover track, muxed over the silent export
python3 scripts/voice-film.py narration.json --film film_silent.mp4 -o film_voiced.mp4
# Optional: background music under the voice — any audio file; it loops/trims to the film,# fades in/out, and sidechain-ducks ~10dB while the voice speaks (--no-duck to disable).# Sourcing the track is upstream of this skill: a music library or a generative# music service. This skill never sources audio; it only places what you give it.
python3 scripts/voice-film.py narration.json --film film_silent.mp4 \
--music bg.mp3 --music-volume 0.06 -o film_voiced.mp4
Rules that make this work (paid for on the Trinity export — details in reference.md →
Voiced export):
Freeze the source first. Copy the film to a working path and hash it before/after the
render. Exporting a file that's mid-edit produced a corrupted render once already.
Narration = captions. In verbatim mode the two are the SAME strings — so a trim to
fix a collision edits the film's caption AND narration.json together, and you re-render the
silent MP4 after any text change (the words are on screen). Parity-check the two arrays
(extract the CAPS strings, diff against narration.json, require 0 mismatches) as part of
verification. A cloned voice runs ~2.0–2.2 words/sec on clause-heavy lines — slower than
you'd guess. Don't pre-trim by guesswork: generate once, read the fit report, trim only the
collisions. The clip cache re-bills only edited lines.
Fix pace before words. Voices differ hugely (ElevenLabs' Daniel reads ~40% slower than
a typical clone). If the first report is a wall of collisions, set "speed" in
narration.json (or --speed, 0.7–1.2) and regenerate — speed compresses pauses too, so
1.15 shortens clips ~28%. Only then trim the survivors by hand.
Collisions are errors; spills are style. Speech overlapping the next line's start is
refused by the tool. Speech running past its own caption's fade is fine.
Verify placement objectively: the fit report proves fit, not placement — after muxing,
run silence-detection over the output and diff speech onsets against narration.json times
(expect two benign artifacts: back-to-back lines merge, and a breath fires ~0.5s early).
Verify ducking by measurement, not by ear: compare volumedetect mean levels of a
speech span against a voiceover gap. Speech span ≈ voice-only level (music adds ≤3dB);
gap shows the music at full background level. A "duck" that measures 0dB is the
sidechain-threshold trap (reference.md → Voiced export → Traps).
Requires: ffmpeg on PATH, playwright resolvable (cwd / $PLAYWRIGHT_DIR),
$ELEVENLABS_API_KEY (or --env /path/to/.env), and a voice id (voice_id in
narration.json, --voice, or $ELEVENLABS_VOICE_ID).
Model: default is eleven_v3 (most expressive tier). v3 ignores
voice_settings.speed — the request returns 200 and the audio is unchanged (measured), so
the "speed" pacing knob only works on eleven_multilingual_v2 (pin it via model_id in
narration.json when you need it). On v3, pace with --max-speedup 1.1 (auto-compresses only
colliding clips, transparently) and trim the survivors by text.
Breathing gap: the fit enforces --min-gap (default 0.3 s) of silence between each
line's speech end and the next line's start. Starts never move — they are the film
alignment; the previous line absorbs the gap via atempo or a trim.
Outputs
A standalone film.html (durable working source — save it out of any scratchpad)
A shipped component reading live data as props
Beat screenshots proving each beat renders
A written note of what is verbatim vs composed, surfaced on the page
(If exported) narration.json + the voiced MP4 + vo_clips/ cache, saved next to the film
source — the narration script is part of the working source, not a throwaway
Completion checklist
Data interrogated; unsupported story beats cut and reported
Palette sampled from the real reference / brand tokens
One real worked example, verbatim, with sourcing disclosed
No un-disclosed scale mixing; no fabricated statements or metrics
The unfinished part is visibly unfinished
Every beat screenshotted and reviewed; every computed value measured
Component reads live data (change the data → the film changes)
Build passes; real page driven; scrubbing verified
Working source saved somewhere durable, not a scratchpad
(If exporting) narration fits: 0 collisions in the fit report; onsets verified against
narration.json after the mux; the stamp's date disclosed wherever the video is posted
(If verbatim mode) captions ⇄ narration parity: 0 mismatches, checked by script — and
every caption trim re-rendered the silent MP4, not just the audio
Reference
reference.md — code recipes: film skeleton, camera + clamping, path-following travellers,
the integrated-rate trap, growth packing, the React port, palette sampling, measuring, and
the voiced-export pipeline (frame render, narration format, pacing traps).
Self-Improvement
After completing this skill's primary task, consider tactical improvements:
Review execution: Were there friction points, unclear steps, or inefficiencies?
New trap found? If a bug cost real time, add it to reference.md → Traps with the
symptom, the cause, and the fix. Symptom-first — that's how it'll be searched for.
Scope check: Only tactical/execution changes—NOT changes to core purpose or goals
Apply improvement (if identified):
Edit this SKILL.md (or reference.md) with the specific improvement
Bump metadata.version and prepend a changelog entry
