Files
hyperframes/skills/pr-to-video/references/story-design.md
T
WaterrrForever 7d21cc9b8a fix(skills,cli): close four reproduced contract gaps from the CLI feedback digest (#2476)
* fix(cli): invalidate the skills nudge cache after a successful install/update/check

The passive "N skills out of date or missing" nudge reads a 24h config
cache that only the background check (on non-skills commands) ever wrote.
The skills commands themselves are excluded from the nudge pipeline, so a
successful `skills update`/install/check never refreshed or dropped the
cached verdict — the pre-install count kept printing on every other
command for up to 24h.

Reconcile commands now drop the cached verdict (counts + timestamp) so
the next command's background check re-runs for real. The offline
presence-only path deliberately keeps the cache: that run learned nothing
about freshness.

* fix(skills): win32-safe npx spawns in media-use + accurate whisper wording

The Whisper transcribe fallback and the Kokoro local-TTS delegation both
spawned a bare "npx" via execFileSync — on Windows npx is npx.cmd, which
spawn cannot exec, so both paths died with `spawnSync npx ENOENT`. Route
them through the skill's existing resolveSpawnCommand (node + npx-cli.js
on win32, no shell:true), same as the audio engine's TTS spawns.

Also corrects the "bundled with the hyperframes CLI" claim about
whisper.cpp: it is resolved from PATH / installed via Homebrew / built
from source with git+cmake on first use, and models download from
HuggingFace — nothing whisper is shipped in the package.

* feat(skills): canonical fully-silent marker + auth status exit-code docs

product-launch's Step 3.1 gate said "or the project is marked silent"
but nothing defined how to mark one, and audio.mjs unconditionally
retrieved BGM. Define the canonical marker — `music: none` in the
storyboard's top YAML block, plus no SCRIPT.md — and honor it:
audio generate produces nothing (removing stale audio_meta.json, since
absence is what assemble treats as silent), and `music: none` with
narration keeps TTS while turning BGM off.

Also documents the `auth status` exit-code contract (exit 1 while
signed out is the normal offline state, not a failure) in the
product-launch Step 0 note and the CLI skill's cloud reference.

* fix(skills): transient-init retry for standalone animation-map and contrast-report

The standalone helpers called initializeSession exactly once, so a valid
modular project — whose sub-composition timelines register asynchronously
— could hit the readiness deadline and die with the transient
"zero duration / Runtime ready: false" diagnostic the render pipeline
retries (probeStage). Add initializeSessionWithRetry to the shared
package-loader (both byte-identical copies): close the crashed session
and retry once with a fresh browser, gated by the engine's canonical
isTransientBrowserError — now re-exported from @hyperframes/producer,
with a frozen fallback pattern list for older published packages. The
"Runtime ready: true" fast-fail (a genuine authoring bug) still fails
without a retry.

* feat(skills): extend the fully-silent marker to faceless-explainer and pr-to-video

Both workflows reuse product-launch's audio model — their Step 3.1 gates
carried the same undefined "marked silent" phrase, and their (intentionally
identical) audio.mjs copies had the same unconditional BGM retrieve. Port
the `music: none` marker handling into both copies, define the marker in
their SKILL.md Step 3.1 and story-design references, and turn the
copies' "intentionally identical" header claim into a byte-identity pin
test so the next fix can't silently miss one of them.

* test(cli): reset the prune mock explicitly instead of relying on restoreAllMocks

The converge test's toHaveBeenCalledTimes(1) held only because vitest 3's
vi.restoreAllMocks() clears vi.fn() call state; vitest 4 restores spies
only, so the count would accumulate across tests and fail. Reset
pruneOrphanedLockEntries in beforeEach like the other manifest mocks —
passes under both vitest 3.2.4 (pinned) and vitest 4.

* test(skills): close review findings — package-loader pin, whisper win32 parity, quoted-none

Review follow-ups on #2476:

- package-loader.mjs byte-identity pin (the elevated concern): the two
  copies now carry initializeSessionWithRetry + FALLBACK_TRANSIENT_PATTERNS,
  exactly the shared-logic shape a future fix could land in one copy and
  miss in the other — same enforcement as the audio.mjs pin.
- whisper win32 call-site parity: runWhisper's npx resolution lifted into
  lib/npx-sync.mjs (resolveNpxInvocation, injectable params matching the
  localTtsGenerate idiom) with the same three-branch coverage as the
  Kokoro site — plus the hard-fail contract (throws actionably, since the
  whisper fallback has no next provider to fall through to).
- quoted music: "none" pin: the vendored storyboard parser strips matching
  quotes at parse time (stripQuotes), so the silent marker already accepts
  the quoted spelling — pinned so that stays true.
2026-07-15 22:22:16 +08:00

28 KiB
Raw Blame History

Story design — PR → narrative

Use this reference in Step 3 to write STORYBOARD.md and SCRIPT.md for a PR-to-video — a code change (the diff, commits, files, +/ stats, and the people behind it) turned into an explainer. There is no website and no captured assets; the PR was ingested into capture/extracted/ in Step 1.

This file defines the story: what the video explains, in what order, and why each frame exists. It does not define layout, effects, animation, or file syntax. For exact storyboard syntax follow ../hyperframes-core/references/storyboard-format.md and ../hyperframes-core/references/script-format.md.

Read first

  1. hyperframes.json — locked brief: angle (archetype), audience, length, aspect, language.
  2. frame.md — tone, type, design system (the shipped preset is claude: warm editorial, a serif that thinks, scarce coral, a navy code surface).
  3. capture/extracted/visible-text.txt — the assembled PR brief: title, meta (base ← head · +N/M across F files), people, body, commits, changed files, and a budget-bounded set of representative diff hunks. This is your source of information.
  4. capture/diff.patch — the full unified diff, for deeper hunk selection than the brief's excerpt.
  5. capture/extracted/people.json — contributors (author / committers / reviewers / commenters), bot-filtered, each with an avatar in assets/<login>.png (for the credits close).

Output

  • STORYBOARD.md — the explanation plan, one frame per beat.
  • SCRIPT.md — the locked narration, only for spoken frames.

Every frame includes the required storyboard-format fields plus the narrative metadata below.

Core rule

A diff is a list of edits. A video is a guided act of understanding.

Do not narrate the diff file-by-file or read the PR description aloud — that is the single most common failure. Explain the change — and where the change has a runtime behavior, show that behavior in motion (a mechanism beat — see "Show the behavior" below), don't just display the lines that changed. Reorder, merge, omit, compress: surface the one change that matters and drop the incidental churn (lockfile bumps, formatting, generated files) unless it is the story. Scene order comes from narrative design, not from the diff's file order or the commit list.

Value before evidence (../hyperframes-creative/references/story-spine.md): the viewer-facing payoff — what the change unlocks, fixes, or speeds up — lands by the second beat; the diff and the mechanism are the evidence for that claim, never the opening. Implementation is the footnote of the story, not the spine.

Default to a plain, technical, unhurried developer voice — accurate, specific, no hype, no marketing gloss. You are explaining a real change to engineers; respect their time and intelligence. frame.md (claude) tunes the voice toward considered and literary; it does not change the structure.

PR archetypes

Choose one archetype (or name a compound). Each is a complete path through understanding a change — do not splice phases from different archetypes.

  • Changelog — "here's what shipped." Hook naming the headline → 24 roughly co-equal change items → ship/wrap. Best for release PRs, multi-change PRs, "what's new in vN." Items are parallel → cut / push-slide between them. Rule-of-three is strongest when changes compress. An item with a visible behavior can be a mechanism mini-demo instead of a bare diff.
  • Feature-reveal — "here's what you can do now." Hook (the outcome the feature unlocks, in user language) → the payoff made concrete (impact — what now works) → name it (change) → the new code typing on (diff) → animate what it does (mechanism) → close (a callback to the promise). Best for a PR that adds one notable feature. The promise leads and the code proves it: diff and mechanism are the evidence for the opening claim — the viewer should already care before the first line of code appears.
  • Fix-explainer — "this was broken; here's the fix." Symptom/bug (problem) → animate the broken behavior (mechanism) → the fix as a before→after (diff) → the behavior now working (mechanism) or the result (impact). Best for bugfix PRs. Seeing the bug happen and then not happen is the turn — a stronger shape (tension → turn → relief) than the diff alone.
  • Refactor-walkthrough — "same behavior, better shape." Hook (the smell / the why) → old shape vs new shape (before_after) → the structure untangling, same inputs → same outputs (mechanism) → payoff (evidence — lines removed, perf delta, files touched). Best for refactors, perf, cleanups, migrations. A mechanism animation proves "same behavior, better shape" far better than asserting it.

Choosing: one notable new capability → feature-reveal; a bug fix → fix-explainer; a behavior-preserving cleanup/perf/migration → refactor-walkthrough; many co-equal changes / a release → changelog. Tie-breakers: a feature that also fixes a bug → feature-reveal with the fix as one body beat; a fix that needed a small refactor → fix-explainer (the fix is the headline). Compound: write arc as "<outer> with <inner>", e.g. "feature-reveal with changelog". Outer = the macro arc the viewer rides; inner = the body rhythm.

PR-native frame types

Set each frame's type to one of these PR-native values. (The storyboard parser keeps type verbatim; it is a narrative + pacing label, not a hard enum.) Each maps to a claude frame treatment and a typical visual — so the type, the design, and the visual stay aligned end to end. Note mechanism is the show-the-behavior beat (an invented animated diagram), distinct from diff (show the code).

type The frame's job claude treatment (frame.md) typical visual (see code-vocabulary.md)
hook The high-leverage opening 35s Cover — (or code-3d-extrude for a hero code moment)
problem The bug / smell / pain / why-care the PR resolves Statement or Pull-quote code-highlight (spotlight the offending line)
change Name the change / the feature / the PR itself Statement or Cover
diff The change body — a before→after, a hunk, new code typed on Code Surface (navy) code-diff / code-morph / code-typing
before_after Explicit old-shape vs new-shape comparison (refactor/fix) Code Surface (split / morph) code-morph / code-diff
mechanism Show what the change DOES at runtime — the request retrying, the cache filling, serial→parallel, the race resolved invented diagram on cream (hairline ink + one coral active marker) invented SVG/GSAP; flowchart / flowchart-vertical / data-chart where they fit
impact The payoff — what now works, what's now possible; opens the video as the promise (feature-reveal) or lands it as the result Number / Impact number-lockup (no code block needed)
evidence Concrete grounding — +N/M, a passing test, a benchmark Number / Impact code-diff red→green / number-lockup
credits Shipped-by close — the humans behind the change Closing — (avatar row from assets/<login>.png)
cta The closing ask — pull it, upgrade, read the PR Closing — (coral-callout)

The body of a PR video alternates diff (show the code that changed) with mechanism (show what it does at runtime), landing on impact / evidence (the result). A body that is all diff reads as code show-and-tell — the mechanism beat is what makes the change legible and is the usual cure for a video that feels flat. Every PR has a change, so at least one diff (or change) frame always exists; most PRs also have a behavior worth animating.

Hook strategy

The hook is the highest-leverage 35 seconds. Pick one:

Strategy When Example
Shocking statistic The change quantifies the stakes "This PR deletes 1,200 lines." / "40% faster cold starts."
Counterintuitive claim The change contradicts intuition "We made the client slower — and that fixed it."
Pain validation The audience already feels the bug "Every deploy, the same flaky timeout."
Concept announcement The change has a name worth landing "Meet retry-with-backoff — flaky networks stop killing your requests."
Before/after teaser The diff is the whole story "One line threw. Now it recovers."
Stakes / consequence The "why care now" is a real cost "This crash hit every user on a flaky network."
Direct address The audience is clearly defined "If you've ever waited on a 5-minute CI run…"

Do not open with a generic repo/company description. Whatever the strategy, the hook speaks the viewer's outcome language (story-spine rule 1) — never file / function / identifier names; numbers only when they carry stakes ("40% faster"), not inventory ("23 files changed").

Clarity / rhetoric technique catalog

Each frame's persuasion is a named technique, not "explain the change." Combine when several are active:

  • Make-concrete — Worked example (one real request/input) · Analogy (backoff as "knock, wait longer, knock again") · Concretization (abstract change → one tangible code line)
  • Reveal-in-order — Progressive disclosure (the diff one line at a time) · Build-up (the simple call, then the edge case) · Signposting ("before… after…")
  • Contrast — Before/after diff · Old shape vs new shape · The bug vs the fix · Two approaches compared
  • Structure — Rule of three (three changes) · Numbered enumeration · Question→answer · Frame-then-fill (state the shape, then the code)
  • Evidence+N/M stat · Passing test / green check · Benchmark / perf delta · Causal chain (request → 5xx → retry → success)
  • Memory & landing — Callback (return to the hook's bug) · Distillation (the change in one line) · Generalization (this fix → the principle)

Emotional beats

beat is one word or a short compound (e.g. "Recognition and relief"). Avoid generic "positive". A PR video rides a comprehension arc:

  • Negative valleyopen the gap (hook/problem): curiosity · frustration · recognition · concern · "ugh, that bug"
  • Pivotorient (change): clarity · orientation · anticipation · focus
  • Buildbuild understanding (diff/before_after/impact/evidence): comprehension · "aha" · confidence · momentum · conviction · relief (for a fix)
  • Resolutionland (credits/cta): satisfaction · resolve · "ship it" · inevitability

Compound beats are often strongest: "Recognition and relief" (a fix), "Curiosity and confidence" (a feature).

The body is a sequence

A PR video's core is 25 body frames, each advancing one change / one before→after / one item, building cumulatively. Alternate diff (the code) with mechanism (the behavior) — don't stack code surfaces:

  • changelog: a diff (or a mechanism mini-demo) per change item; parallel → default cut / push-slide.
  • feature-reveal: impact (the promise, concrete) → change (name it) → diff (the code, often typing/morphing on) → mechanism (animate it working) → a closing callback to the promise.
  • fix-explainer: problem (symptom) → mechanism (the bug happening) → diff (cause + fix, before→after) → impact (result, or a mechanism of it working).
  • refactor-walkthrough: before_after structure → mechanism (the structure untangling, behavior preserved) → an evidence numbers beat.

Continuity across frames

This framework builds one frame per worker — there is no multi-frame "continue run." A sequence reads as one continuous shot through two storyboard-level levers, both yours:

  1. A consistent stage — consecutive body frames share one composition idea (the same navy code window filling in, the same before|after split, the same counter advancing), stated in each frame's scene so Step 4 and the workers keep the stage stable.
  2. A consistent transition — pick one seam type for a run (crossfade for a soft code reveal, push-slide for the next change item) and repeat it.

When a single element genuinely transforms between two ideas (the failing test flips green, the old function becomes the new one), keep it within one frame as a development beat (entrance → transform → settle) — the worker owns that motion (a code-diff or code-morph block). Note the intent in scene / narrative; Step 4 wraps it in a time-coded shot sequence around the block (a code-diff / code-morph).

Transitions

Use only registry transition names in transition_in:

cut | crossfade | blur-crossfade | push-slide LEFT | push-slide RIGHT | push-slide UP | push-slide DOWN | zoom-through | squeeze

Pick 23 for the whole video and repeat. Frame 1 is cut (no previous frame). Match the seam to the narrative: ordered change items → a consistent push-slide; a soft reveal / into-the-cause → crossfade / blur-crossfade; zooming into a code line or pulling back to the file tree → zoom-through; a clean new change item → cut.

The diff is the centerpiece

Code beats live on the navy code surface (claude's Code Surface treatment) — but the body is not all code (pair them with mechanism beats, next section). Plan the code beats deliberately:

  • Feature 24 real diff hunks, named in each frame's scene — each a small, legible snippet (~412 lines), never a whole file. Pull them from capture/diff.patch / the brief's "Representative diff."
  • Name which code animation block the frame wants in scene (the Step-4 visual phase and the worker read it). See code-vocabulary.md for the full map; the short version: before→after = code-diff; refactor/rename continuity = code-morph; new code written on = code-typing; spotlight one line = code-highlight; walk a long file = code-scroll; a hero reveal = code-3d-extrude / code-particle-assemble.
  • Numbers (+1,204 / 318, files touched, perf delta) belong on an impact / evidence frame as a number-lockup, not read aloud in narration.

Show the behavior — the mechanism beat (not just the diff)

A diff shows what changed in the code. It does not show what the change does — and "what it does" is usually the more memorable, more explanatory beat. The single biggest reason a PR video feels flat is that every body frame is a code surface or a number: it tells (here are the lines, here is the stat) but never shows (here is the request actually recovering).

A mechanism frame animates the runtime behavior the PR changes — built as an invented animated diagram (SVG / HTML / GSAP on claude's cream ground: hairline-ink nodes / edges / lanes, one coral marker on the active or changed element), where the build is the teaching — each part appears on beat, the flow plays out across the shot. It is not a code block and not a headline. Reach for the flowchart / flowchart-vertical / data-chart registry blocks where they fit; otherwise invent it (visual-design.md's diagram / abstract-graphics register).

Plan at least one mechanism beat for any PR with a visible runtime behavior (most feature and fix PRs have one). What to animate, by what the change touches:

The change touches… Animate (the behavior, not the code)
Retry / backoff / resilience a request lifecycle: fire → 500 → wait (delay growing) → retry → 200
Caching / memoization two lanes racing: cold (slow, hits the DB) vs cached (fast, hits the cache)
Concurrency / parallelism a serial single lane reshaping into parallel lanes
Race / ordering bug the broken behavior first (items dropped, two writers colliding), then the fixed flow
Performance two timelines / bars racing, the new one finishing first (a data-chart fits)
Refactor / migration a tangled call-graph untangling into a clean one; same inputs → same outputs
New endpoint / pipeline / state data flowing through the new path; a state machine lighting up step by step (flowchart-vertical)

Name the mechanism in the frame's scene ("animate the request retrying: fire → 500 → backoff → 200, invented SVG flow") so Step 4 and the worker build it. The diff frame and the mechanism frame are complementary — the diff is the proof in code, the mechanism is the proof in motion; alternate them rather than stacking code surfaces.

The close: a credits / shipped-by scene

A PR is shipped by people, and every PR video closes with a credits frame naming them. capture/extracted/people.json lists real contributors (bot-filtered), and Step 1 downloaded each avatar to assets/<login>.png (the avatarFetched: true entries — confirm with ls assets/). reviewDecision (e.g. APPROVED) is honest grounding.

The PR author only opened the PR — not necessarily who wrote the code. A teammate often authors most commits. Lead the credits with committers by commitCount, not the opener.

The credits frame is an avatar row with names + roles + an "approved" check. On that frame only, set asset_candidates to 16 entries of assets/<login>.png — <login>, <role> (commit authors by commitCount first, then reviewers; only avatarFetched: true logins). The body stays code-only — avatars appear only on this close, never decorating a diff frame. The frame sits in the Step 3 proposal like any other, so the user can cut it there; skip it yourself only when no avatar was fetched.

Narrate the name, not the handle. people.json carries a name field (GitHub display name, e.g. "Miguel Angel Simon Sierra") next to login for whichever contributors gh already named (author, commit authors, mergedBy); it's null for reviewers/commenters/assignees, which gh pr view only ever gives a bare login. Before writing this frame, resolve any null name yourself for the 1-6 people going on the close: gh api users/<login> --jq .name. Voiceover always says the name (first name is enough — "Shipped by Miguel, reviewed by Wenbo") and never reads a raw @login aloud (@miguAng18947550 spoken by TTS is the failure mode this exists to avoid). On-screen text under each avatar can show both, name first, handle small and secondary: Miguel Angel Simon Sierra / @miguel-heygen. When a name still doesn't resolve (GitHub has no public name for that user either), fall back to the login on-screen and skip that person from the spoken line rather than reading the handle.

Every other frame has no asset_candidates (the visuals are invented downstream from scene + the diff).

Versions on the end card (cta / changelog)

A cta ("upgrade to vN", "npm i pkg@N") or a changelog "what's new in vN" wants a real version — and a version is the one fact you must never invent. A PR carries no shipping version, so Step 1 resolves a best-effort one for MERGED PRs and writes it into capture/extracted/visible-text.txt as a Shipped in: <version> (<source>) meta line (mirrored in capture/pr.json as shipped_version / version_source). Use it:

  • Shipped in: present → use that exact version on the end card. A version_source of unreleased means the change is on the default branch but not yet in a tagged release — say "shipping in the next release" rather than pinning a tag.
  • No Shipped in: line (open PR, or no version resolvable) → state the repo / PR URL only ("read the PR at github.com/…", "pull it") and do not name or guess a version number.

Per-frame length budget — ≤ 9 s, word count is the real measurement

The largest quality bug in PR videos is scripts that talk too long. TTS runs at ~2.2 words/second, so a 45-word "7-second" script is really 20 seconds, and the visual phase has to pad the tail with idle drift (the video reads as "shimmering"). Budget by word count:

Bound Words (@2.2 wps) Duration When
Soft target — default ≤ 19 ≤ 9 s Every frame aims here; the cut stays alive.
Exception — ≤ 2 frames ≤ 26 ≤ 12 s The main diff (the one change you must explain) or a causal-chain change.
Hard cap > 26 > 12 s Trim or split.
Whole-film target ≤ ~400 up to ~3 min Sweet spot ~3090 s (≤ ~155 words); the body carries the load.

Estimate while writing: duration ≈ ceil(word_count / 2.2). 29 words → 13s (trim); 17 words → 8s; 12 words → 6s. Trim techniques: cut the lead-in clause ("Until now, the agent shipped…" → "The agent shipped…"); move numbers off-script onto a counter; split only when the halves carry distinct beats (cause then effect). Silent frames are allowed and common — a diff typing on, a before→after morph, a counter running. Set voiceover empty, omit from SCRIPT.md, and make narrativeRole carry it. A complex change does not need a long script; it needs a careful one — if you can't headline the change in 19 words, the headline isn't sharp yet.

Write each line as discrete cues, not one run-on breath. Step 5 reveals each on-screen piece when the voiceover names it (the anti-PowerPoint mechanism). A line with clear phrase boundaries — "Three retries — then it backs off — then it gives up clean" — hands the shot its reveal cadence for free; a single long clause leaves the frame nothing to pace to.

Music & silence

The storyboard's top YAML block carries a music: field — the BGM mood the audio step retrieves against (e.g. music: confident minimal tech underscore). Omitting it falls back to message:arc: → a neutral default, so BGM plays unless turned off explicitly.

  • music: none — BGM off (narration, if any, still runs).
  • music: none + no SCRIPT.md — the canonical fully-silent marker: no narration, no BGM, no SFX. audio.mjs generates nothing and the audio step is a clean skip. Use exactly this spelling when the user asks for a silent / music-free video.

Frame template

## Frame N — Short name

- scene: one clear visual idea — name the hunk/file + the code-\* block ("the request() retry block, ~6 lines, code-diff")
- voiceover: "spoken guide text, or empty"
- duration: ceil(word_count / 2.2) seconds
- transition_in: crossfade
- status: outline
- src: compositions/frames/NN-short-name.html
- type: diff
- persuasion: Before/after contrast
- beat: comprehension
- blueprint: <candidate id from the role→blueprint menu, or omit — a code beat usually omits it (the code-\* block is the shape)>

narrativeRole: What this frame does in the viewer's understanding (its job, not what's on screen).
keyMessage: The one thing the viewer should understand after this frame (one sentence).

The credits frame additionally carries an asset_candidates: line (see the credits section); no other frame does.

Final checklist

  • One archetype is named (compound only when explicit); the sequence is narrative-driven, not diff-order-driven.
  • The opening uses a named hook strategy; you do not read the PR description aloud.
  • The hook is in viewer-outcome language (no file / function / identifier names), and the video's message lands by beat 2 (story-spine).
  • Each frame has one job; the body builds cumulatively, alternating diff (the code) with mechanism (the behavior) + impact / evidence — not a single isolated body frame, and not an unbroken stack of code surfaces.
  • Every frame has type (PR-native), persuasion (a named technique), and beat (specific). The emotional arc matches the archetype (fix = frustration → relief; feature = curiosity → confidence).
  • Each voiceover is phrase-segmented into cues (each a piece Step 5 can reveal on), not one run-on clause; a candidate blueprint: is tagged from the role→blueprint menu where a proven shape fits (a code beat usually omits it — the code-* block is the shape).
  • 24 real diff hunks featured, each a small legible snippet (not a whole file), each naming its code-* block in scene.
  • At least one mechanism beat animates what the change does at runtime (an invented diagram, or a flowchart / data-chart), named in its scene — unless the PR genuinely has no visible behavior (a pure docs / config bump). The body is not an unbroken run of code surfaces.
  • Transitions use only registry names and repeat 23 types; frame 1 is cut.
  • The video closes with a credits frame (skipped only when no avatar was fetched); asset_candidates is absent on every other frame (16 assets/<login>.png entries on the close, avatarFetched: true only).
  • Each script fits the budget — ≤ 19 words / ≤ 9 s default, ≤ 2 frames at the ≤ 26 / ≤ 12 s exception; duration = ceil(word_count / 2.2), not a guess.
  • SCRIPT.md contains only locked spoken narration; silent frames are intentional and omitted from it.