Skip to content

Commit 370f886

Browse files
fix(skills): style-sampling bug bash fixes, and the five style packs re-pinned (heygen-com#7164)
* fix(skills): style-sampling follows what the kits can do Stop 1 poses on the kit's free timing estimate instead of a plain page, a voice id prefix is resolved from the account's clones, speed is measured as REMIX says and changed locally, Pexels without a key and sites that refuse agents move to the next source, and stop 2's preview opens no browser. Co-Authored-By: Claude Opus 5.5 <[email protected]> * fix(skills): style-sampling answers the second bug bash A stock voice is made only at the voice step, after the storyboard; an unfit topic gets the nearest angle in the first reply; credits name only who the person names; logos and page screenshots need terms that allow them; a long render gets a time estimate and progress; and picking a finished episode back up loads this skill. Co-Authored-By: Claude Opus 5.5 <[email protected]> * chore(cli): pin the five style packs with both bug bashes fixed ffmpeg 9 works, a voice line is bought once and re-timed for free, a crash after a paid call never pays twice, the run's credit limit holds in every kit tool, the storyboard poses on free estimated timing, checks pass on the shipped components, and pulled files carry a real date. Co-Authored-By: Claude Opus 5.5 <[email protected]> * fix(skills): keep the style-sampling description within the 320-byte budget Co-Authored-By: Claude Opus 5.5 <[email protected]> * fix(skills): a stock voice is picked by language with --plan, and resuming routes through the project's CLAUDE.md media-use resolve --type voice needs --intent and returns an audio asset, not a voice id the kit can reuse. generate --plan names the first public voice in a language without making anything. The description goes back to main's wording; a finished project's CLAUDE.md now says changes go through style-sampling instead. Co-Authored-By: Claude Opus 5.5 <[email protected]> * fix(skills): style-sampling offers a free Pexels key in the one round of questions Co-Authored-By: Claude Opus 5.5 <[email protected]> * docs(skills): style-sampling thanks the creators these styles learned from, without naming them Co-Authored-By: Claude Opus 5.5 <[email protected]> * fix(skills): style-sampling lists a presenter's looks with media-use avatars, never the heygen sign-in Co-Authored-By: Claude Opus 5.5 <[email protected]> * chore(cli): pin the five style packs with the plain end card and interruption-safe paid takes Co-Authored-By: Claude Opus 5.5 <[email protected]> * docs(skills): style-sampling names the creators these styles learned from, and that they don't endorse them Co-Authored-By: Claude Opus 5.5 <[email protected]> * docs(skills): trim the style-sampling thanks line Co-Authored-By: Claude Opus 5.5 <[email protected]> * chore(cli): pin the style packs with the restart fixes from review Co-Authored-By: Claude Opus 5.5 <[email protected]> --------- Co-authored-by: Claude Opus 5.5 <[email protected]>
1 parent 71b67cc commit 370f886

4 files changed

Lines changed: 27 additions & 22 deletions

File tree

‎packages/cli/src/commands/templates.ts‎

Lines changed: 10 additions & 10 deletions
Original file line numberDiff line numberDiff line change
@@ -31,28 +31,28 @@ export interface TemplatePack {
3131
export const TEMPLATES: Record<string, TemplatePack> = {
3232
"animated-science-explainer": {
3333
id: "M9003",
34-
sha256: "4ead405395892545c658a03b48630529943f5598cc15770930c0ffa6b8de58e9",
35-
bytes: 25559788,
34+
sha256: "42fc622292e323c5d620480428a4f3162b6a8844c0330abf29a367fa214bbe6a",
35+
bytes: 25535207,
3636
},
3737
"finance-explainer": {
3838
id: "M9002",
39-
sha256: "b10efcfad35097ddac8118062701eeb3ef4012f3c9b41704e474964d09d37d11",
40-
bytes: 140688140,
39+
sha256: "52be65859022d8c6b1ed2e3701d875d95b6dce87a8e5bf16e4f2581a56cf746e",
40+
bytes: 140483725,
4141
},
4242
"3d-true-crime": {
4343
id: "M9005",
44-
sha256: "53d1f51eb5b78e3cb60a23625d160952236d5f5fb8d41118798d782a6168eec1",
45-
bytes: 5977486,
44+
sha256: "62959b0b978fb97727f06346b9ea5aaf3d8dcf6992f574cbe525bbe893cff287",
45+
bytes: 6544603,
4646
},
4747
"collage-explainer": {
4848
id: "M9001",
49-
sha256: "395362e3c629706143724e99d72dacb4997d35b978ecc4173901cdef6e70205c",
50-
bytes: 36711101,
49+
sha256: "a25529d8ee1dfb09187a6ccc0978d7127a4306992287e68c06f358914d9f1496",
50+
bytes: 36702236,
5151
},
5252
"3d-brand-explainer": {
5353
id: "M9004",
54-
sha256: "4e5c5eb574f977d3f97a170a5a0d8d442234d34dd51f925dc4a2b93e32826ae8",
55-
bytes: 20181051,
54+
sha256: "24d1f3e4bc1d5e6d843e12fc094223f8df6f27f798094940f23c33076ac434a6",
55+
bytes: 20126746,
5656
},
5757
};
5858

‎skills-manifest.json‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -34,7 +34,7 @@
3434
"files": 2
3535
},
3636
"style-sampling": {
37-
"hash": "99809938d12ad2c4",
37+
"hash": "0670f6826ea02c49",
3838
"files": 2
3939
}
4040
}

‎skills/style-sampling/SKILL.md‎

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -9,6 +9,8 @@ description: Make a new episode in a proven video style - Collage Explainer, Fin
99

1010
A good episode is one the series' own viewers would take for the next upload: its structure, pacing, look, graphics and sound, on a new topic, in the person's voice and, when the series has a presenter, their own face. Never use a creator's face, voice, footage, logo or show name.
1111

12+
Thanks to the creators whose channels these styles learned from: Vox (America, Explained) for the Collage Explainer, 小Lin说 (一口气了解) for the Finance Explainer, Kurzgesagt for the Animated Science Explainer, and Fern for the 3D Brand Explainer and 3D True Crime. Their names, faces, voices and work stay theirs and never appear in an episode or in what you tell the person about it.
13+
1214
Style templates are a HyperFrames Pro feature. The CLI pulls them and the server checks the plan.
1315

1416
| Template | Slug | Series | Makes | The person brings |

‎skills/style-sampling/references/remix.md‎

Lines changed: 14 additions & 11 deletions
Original file line numberDiff line numberDiff line change
@@ -6,56 +6,59 @@ Order: `read → settle the voice → say what's coming → ask (only what is op
66

77
## 1. The template
88

9-
- It is in `templates/<slug>/`: CHANNEL.md (the channel: identity, caption skin, title look, rights) and one `series/<series>/` folder with TEMPLATE.md, REMIX.md, voice.md, frame.md, editing.md, footage.md, sound.md, `components/` and `kit/` (the starter project and its tools, with its fonts, sound and assets). A drawn series adds STYLE.md and `kit/engine/`.
9+
- It is in `templates/<slug>/`: CHANNEL.md (the channel: identity, caption skin, title look, rights) and one `series/<series>/` folder with TEMPLATE.md, REMIX.md, voice.md, frame.md, editing.md, footage.md, sound.md and `kit/` (the starter project and its tools, with its fonts, sound and assets). A series built from components has `components/`; a drawn series has STYLE.md and `kit/engine/` instead.
1010
- Where a template names something only the HyperFrames app has (`media.find`, `voice.cloneVoice`, the Storyboard tab), use its command-line route or what this file says.
1111

1212
## 2. Read in this order
1313

1414
1. CHANNEL.md.
15-
2. `series/<series>/TEMPLATE.md`: what the episode is, its runtime, the slots, what topics it is good and not good for. A topic it is not good for: say so in one line and offer the nearest angle that fits.
15+
2. `series/<series>/TEMPLATE.md`: what the episode is, its runtime, the slots, what topics it is good and not good for. A topic it is not good for: say so in one line and, in the same message, offer the nearest angle that fits. Never offer to make the unfit topic anyway.
1616
3. `series/<series>/REMIX.md`. **Its step table is the process.** Follow it in order and skip nothing.
1717
4. `series/<series>/kit/README.md`: the tools, their Python packages and the project they build.
1818
5. Every other file when a REMIX.md step points to it. editing.md lists every component with the creator's own use of it; footage.md says what to show for each kind of line and where to find it.
1919

2020
## 3. What to ask, once
2121

22-
The person's message carries what it carries. Ask once, together, only what is still open: the topic, the length, the voice and, for a presented series, the face.
22+
The person's message carries what it carries. Ask once, together, only what is still open: the topic, the length, the voice, for a presented series the face and, when no `PEXELS_API_KEY` is set and the template's footage.md uses Pexels, whether they'd add a free Pexels API key (pexels.com/api, a minute to get). It is optional: without it, footage comes from the other sources footage.md names.
2323

24-
- **Never the creator's face, voice, footage, wordmark or show name.**
25-
- **A presenter runs on the plan.** Their own HeyGen avatar if they have one, otherwise Avatar IV from their photo, both through `npx hyperframes media-use generate --type avatar` with the finished voice, so the lips land on it. HeyGen Video 1 is the fallback when neither works.
26-
- **Settle the voice before any research.** A HeyGen voice id the person gave is the voice: use it and ask nothing. Anything else: run `npx hyperframes media-use voices`, which lists the clones on the HyperFrames account making the film; never `heygen voice list`, a separate sign-in that can be another account. One clone: use it. Several: ask which. None, or not signed in: ask once for their voice id, a clone they make in HeyGen from 30 s to 3 min of their own reading, or their own recorded narration; with none of these, say in one line that you'll use a HeyGen stock voice (`npx hyperframes media-use resolve --type voice`), name it in the credits as AI-generated, and go on. Nothing else starts until this is settled.
24+
- **Never the creator's face, voice, footage, wordmark or show name.** That rule is about the creator: the person's own clone, their own recording or a credited stock voice are all fine.
25+
- **A presenter runs on the plan.** Their own HeyGen avatar if they have one (`npx hyperframes media-use avatars` lists the looks on the HyperFrames account making the film; never `heygen avatar list`, a separate sign-in that can be another account), otherwise Avatar IV from their photo, both through `npx hyperframes media-use generate --type avatar` with the finished voice, so the lips land on it. HeyGen Video 1 is the fallback when neither works.
26+
- **Settle the voice before any research.** A HeyGen voice id the person gave is the voice: use it and ask nothing. Only part of one (an 8-character prefix): find the full id in `npx hyperframes media-use voices` and ask only when no clone or several match. Anything else: run `npx hyperframes media-use voices`, which lists the clones on the HyperFrames account making the film; never `heygen voice list`, a separate sign-in that can be another account. One clone: use it. Several: ask which. None, or not signed in: ask once for their voice id, a clone they make in HeyGen from 30 s to 3 min of their own reading, or their own recorded narration; with none of these, say in one line that you'll use a HeyGen stock voice, name it in the credits as AI-generated, and go on. Nothing is voiced for it before stop 1: at the voice step, `npx hyperframes media-use generate --type voice --voice <the script's language> --script "<first line>" --out .media/pick.mp3 --plan` names HeyGen's first public voice in that language without making anything; set the kit's voice id to its `voice_id` and credit that voice by name. Nothing else starts until this is settled.
2727
- **Then say what's coming** in one short message that is not a question, and go on without waiting: the working title and hook, the length, and roughly how long it takes. No prices.
2828
- **The length is the person's, and it beats the template's runtime.** Offer about 1 minute first: the cold open, the title beat and one chapter, the end card short. 2 to 4 minutes: REMIX.md's short cut. A full episode: the series' own runtime. Their own words ("90 seconds", "5 min"): that length. No length answered and none in the topic: about 1 minute, and offer the longer cut in the reply. Scale everything to it: script words (the template's script stats at that runtime), facts, sections, music cues, review rounds (one for a minute, two for a full episode); never pad or trim to reach the series' runtime instead.
29+
- **Credits name only who the person names.** Roles and names come from the person, never from a voice list, an account or a placeholder; a role nobody named is left out.
2930
- Everything else REMIX.md §0 asks (chapters, language) takes REMIX.md's default; write each one, and the length, with why, in the project's `CLAUDE.md`.
3031

3132
## 4. Facts and script, then stop 1
3233

3334
- Facts first, per REMIX.md: every number, date and quote with its source. A research sub-agent is fine.
3435
- Write the script to voice.md and the series' script anatomy; run the script checks REMIX.md names.
35-
- **Stop 1:** pose the script and the shot plan as stills on the board, written as the hyperframes skill's `references/storyboard.md` says: one still per beat, its line under it, what each shot shows and which component makes it. A series kit builds its film from the voice's timing, so never voice to get a board: pose the beats on a plain page with the real components, words and fonts, and give each its estimated seconds. Point the person to `boards/storyboard.html` and wait for their go.
36+
- **Stop 1:** pose the script and the shot plan as stills on the board, written as the hyperframes skill's `references/storyboard.md` says: one still per beat, its line under it, what each shot shows and which component makes it. Never pay for a voice to get a board. A series kit builds its film from the voice's timing, so first give it estimated timing with the free route REMIX.md names for stop 1 (most kits: `python3 tools/voice.py --silent`, silent stand-ins at the series' pace), assemble, and pose in the `index.html` that builds, as `storyboard.md` says. A drawn series poses in the test shot its REMIX.md names instead. Every time on the board is an estimate until the voice exists; say so. Point the person to `boards/storyboard.html` and wait for their go.
3637

3738
## 5. Voice and face
3839

3940
- Voice: the person's voice id, through the kit's voice tool, as REMIX.md says. Then the kit's timing and pause tools. The voice is final before any presenter.
40-
- **Audition before the whole script:** one paragraph (the kit's voice tool, as REMIX.md says), measure speech-only words a minute, and set the speed that lands on the series' rate. An instant clone takes no speed: change the words. Voice the script once; re-voice only the paragraphs that changed.
41+
- **Audition before the whole script:** one paragraph (the kit's voice tool, as REMIX.md says), measure words a minute the way REMIX.md §0 measures them, and set the kit's `voice.speed` to land on the series' rate. The voice call itself takes no speed for an instant clone; the kit changes tempo locally from the kept native audio, so a new speed costs nothing. Past the range REMIX.md allows, change the words instead. Voice the script once; re-voice only the paragraphs that changed.
4142
- Face (presented series): REMIX.md's host step, on the final voice: their own avatar, Avatar IV from their photo, or the HeyGen Video 1 fallback. Their own takes: cut them per editing.md's presenter rules.
4243
- A changed line after the presenter take exists means the whole take is made again, or the lips drift.
4344

4445
## 6. Footage, shot plan, build
4546

46-
- Find footage per footage.md, from the sources it names (Wikimedia Commons, Openverse, NARA, Library of Congress, Prelinger, Pexels, Mixkit, official media). Download with curl. One sub-agent per chapter is fine. Every file gets a licence line in the record REMIX.md names; credit what needs credit.
47+
- Find footage per footage.md, from the sources it names (Wikimedia Commons, Openverse, NARA, Library of Congress, Prelinger, Pexels, Mixkit, official media). Download with curl. Pexels needs the person's free `PEXELS_API_KEY`; its web pages refuse an agent, so without a key go on to the next source. Check each clip's own licence (some Mixkit clips are Restricted). A site that refuses an agent (403, 429, a bot check): move to another source; never work around the check. One sub-agent per chapter is fine. Every file gets a licence line in the record REMIX.md names; credit what needs credit.
4748
- Never take footage from YouTube, TikTok, Instagram or X unless its licence is stated.
49+
- A real brand's logo or sign, and screenshots of web pages, appear only under terms that allow it (the template's rights rules, a press kit's terms, a stated licence); otherwise crop it out, cite the page as a source line, or leave it out, and say so in the reply. Where REMIX.md asks to search YouTube first, that is research: a web search for what the topic's existing videos cover, nothing downloaded.
4850
- Write the shot plan and assemble with the kit's tools, as REMIX.md says. **Before any graphic, find the creator's precedent for that kind of moment in editing.md; with none, don't add it.** A component the template lacks: use the closest one and note it in GAPS.md.
4951
- Keep `templates/` as pulled.
5052

5153
## 7. Checks, stop 2, render
5254

5355
- One heavy job at a time (renders, long ffmpeg, transcription): renders with `--workers 1`, ffmpeg with `-threads 2 -nostdin`. For a change, render only that window.
56+
- Before a render likely to run past 20 minutes (3D at full quality, a long cut), say roughly how long it takes; while it runs, report its progress from the log rather than only that it has not finished.
5457
- Run REMIX.md's checks. Report each of our numbers beside the series' own: cuts a minute, pause density, longest hold, loudness, sound effects on picture changes, the script fingerprint. Name every check that failed.
55-
- **Stop 2:** ask once on the draft: a contact sheet and `npx hyperframes preview`, with what is still a stand-in.
58+
- **Stop 2:** ask once on the draft: a contact sheet and `npx hyperframes preview --background --no-open` (add `--port <free port>` when 3002 is taken), with what is still a stand-in.
5659
- Then the one full render REMIX.md's delivery calls for. A render over 10 minutes starts detached and is watched through its log, because a render stops when the shell that started it exits: `bash -c 'set -m; HYPERFRAMES_RENDER_DETACHED=1 npx hyperframes render -o <out> < /dev/null > render.log 2>&1 &'`.
5760

5861
## 8. Reply and hand back
5962

60-
- Say what the episode is, its length (and, for a short cut, that the full episode is one ask away), every stand-in, every check that failed and why, and the footage licence table. Update the project's `CLAUDE.md` with how it was made.
63+
- Say what the episode is, its length (and, for a short cut, that the full episode is one ask away), every stand-in, every check that failed and why, and the footage licence table. Update the project's `CLAUDE.md` with how it was made, and open it with a line that any change to this episode goes through the style-sampling skill, so a later session runs its sign-in check and asks whether the credit limit still applies.
6164
- Keep LOG.md (each step and how long, no prices or credit numbers) and GAPS.md (where the template was missing, unclear or wrong, with file and section) in the project root. GAPS.md is how the template gets better: never skip it.

0 commit comments

Comments
 (0)