the-catch-staging
Skill2 starsChanged 15 days agoMITView on GitHub
Turn a "The Catch" script into a storyboard and StickStage explainer skit in the night-shift world: Nickel the raccoon (ember hoodie, headlamp), the human lead, data as objects in the world (split-flap boards, coin stacks, neon lines, stamps, evidence boards), camera, music and SFX cues, and the variety rules. Use after the-catch-story, or when asked to storyboard, stage, restage or render a The Catch case.
In this repository13 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- llms.txtllms.txt
- crew-episodeSkill
- explainerSkill
- science-shortSkill
- shorts-editorSkill
- skit-directorSkill
- story-episodeSkill
- the-catch-storySkill
- toon-characterSkill
- topic-to-videoSkill
- whos-wrongSkill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 3 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 8
Look bible: ../planning/2026-10-30-the-catch/frames/ (01–05 world, nickel-01..04 host) and plan §4–§6.
- Distinctive: line 120
No brand anywhere on a cover.
- Distinctive: line 133
Check frame 0 of both cuts and both cover PNGs before any render is handed back.
- Distinctive: line 137
Every 16:9 case also ships a ≤ 30 s 9:16 cut for the Thin Lines Theory channel that sends viewers to the long video: the hook, one clue …
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
The Catch: staging
Look bible: ../planning/2026-10-30-the-catch/frames/ (01–05 world, nickel-01..04 host) and
plan §4–§6. Engine reference: skills/explainer/SKILL.md, docs/explainer.md, docs/figures.md.
Renders run on local Tamtree (no Lambda 15 min limit), so long cases are fine.
World rules
- Night everywhere except the final callback, which may be the first daylight of the video.
- Palette: Night
#0E1A24ground, Teal surfaces, Paper words, one Sodium hero per frame (the answer, the light), at most one Ember element (loss, alarm, stamp), Neon only for data lines. - Fonts: Bricolage Grotesque (key phrases), Big Shoulders Display (numbers), IBM Plex Mono (labels, sources).
- Subtitles on, both cuts (owner 2026-11-01: people watch muted):
"overlay": {"subtitles": true, "captions": {...}}, the stick-story word-timed style. Long cut{"y": 0.73, "fontSize": 72}(clear of the source credit); short{"y": 0.42, "fontSize": 76}(mid-screen, above the Shorts UI). Key-phrase cards (2–6 words) sit in the top band,at(0.33, 0.27, …), never in the subtitle band; drop a card rather than stack it on a figure. - No real brands in the picture (owner 2026-10-31): no logos, wordmarks, mascots, signature colours, or cup, packaging, card or store designs; everything is drawn in the tokens above, with generic labels (STORE, BANK, CARD). Be extra careful with props on screen while the narrator names a company (Case 001's cup once had a green band and a round badge, too close to the named brand). A company name appears on screen only in the source credit. Titles, thumbnails, descriptions and hashtags don't name it; the description's source credit may.
Data as objects (never a chart on a plain fill)
| Need | Object | Kind |
| One big number | Split-flap board that clacks to the value | flap |
| Compare 3–6 values | Line-up: coin stacks / pallets / cups against a height wall | stacks (planned) / prop cups today |
| Growth over time | Neon tube on a wall drawing on | line variant neon (planned) |
| Share of people | Crowd of 100 silhouettes, some lit | crowd |
| A rule / verdict | Rubber stamp slam | stamp |
| Where / a list of rules | Evidence board: pinned cards + red string | evidence (on set evidence-board) |
| A thing in the story | Hero prop under a light cone | prop (drawer, envelope, giftcard, till, ledger, jar, cups, coins, vault, cup, lostbox, book) |
| Words on screen | Key-phrase card, 2–6 words | phrase on layer: "screen" |
| Advice to act on | A phone note typed and ticked | phone |
Engine reference for all of these: docs/explainer.md § The night-shift look. Mark kinds that
don't exist yet as planned in the storyboard so the build list stays honest.
Nickel
- Night Shift design (
frames/nickel-01.pngB): ember hoodie, ears through the hood, headlamp. In the engine he is thenickelfigure kind and speaks as an off-screen voice (voices);risepeeks him up over a counter,beam/beamLengthaim the headlamp. - On screen in ~30% of shots; the headlamp beam is how reveals are lit (the beam lands, the number flips).
- Expressions:
smirk,sly,shock,grin,gotcha,shh. One reaction per reveal, never two in a row with the same face. - Nickel never stands next to a chart pointing at it like a presenter. Nickel sneaks, opens, finds, reacts.
Human lead
- Toon tier (
skills/toon-character/SKILL.md). Appears in the cold open, the human turn and the callback. Their object (the card) is the transition between their world and the money world (match cut: card in the drawer → card on the vault floor).
Motion and sound per beat
- A visual change every ≤ 4 s; camera always drifting; a push-in on every reveal; whip-pan or match cut between scenes.
- Each cue names its SFX: flap
clackper tile, coinclink, stampthud, scenewhoosh, money revealregister, drawerslide. Music:mystery(open, clues),build(last clue),dropon the catch line,resolve(verdict). Duck under the voice. - Voice pace:
fast(~190 wpm). Narrator linespace: "fast"; Nickel a separate, short, dry voice. - Camera: keep the night finish's slow breathing zoom and the push-ins; they must glide (camera scale is written at 4 decimals since f7fe7c1; never round camera values to 0.01, it reads as dropped frames).
Voices (owner decision 2026-10-31: Google TTS)
Gemini TTS google/gemini-3.8-flash-lite-tts through local Tamtree, never say, Chatterbox or
Kokoro in a final render (say only as a timing stand-in while staging).
| Speaker | Voice |
| Narrator | Charon |
| Nickel | Algenib |
| Human lead (Maya) | Leda (pick another female/male Gemini voice for a new lead and record it in the case folder) |
Recipe: local flow catch-voice, input {model, lines:[{id, speaker, text, tts_voice}]}, one item per
beat (text = the beat's spoken if set, else line). Write public/skits/<id>/voice/<beat>.wav and
voice.json (text = the skit line, spoken = what was said, durationMs). Full steps and Tamtree
branch caveats: ../planning/2026-10-30-the-catch/handover.md § Next 2. Gemini voices take ~20% longer
than say on short lines (Case 001 went 3:30 → 4:15), so budget runtime for that.
Approved format (Case 001, owner 2026-10-31: "really well done")
out/the-drawer.mp4 (4:15) and out/the-drawer-short.mp4 (29.5 s) are the reference. Reuse, don't
reinvent:
- Start from the generator. Copy
../planning/2026-10-30-the-catch/cases/001-the-drawer/build_skits.pyinto the new case folder and replace scenes, datasets and lines; keep the helpers (fig,beat,show/hide/setp/focus/sfx,phrase,num,NICK,PEEK,duck) and both cuts in one file. - Same structure: ~10 scenes over the night sets, one chapter per scene, music bed per act,
whipbetween scenes, five acts tagged (hookby 0:08,rehooks, onereveal, callback). - Nickel at the right moments, not rare. Besides his own lines, he peeks (
PEEK+setp rise 1on the key word,duckafter) on every payoff line: the hook question, each big-number reveal, each stamp, the named mechanism ("breakage"), the human consequence, the rules, and he stands beside the phone for the "tonight" actions. One expression per peek, never the same twice in a row. - Every hidden figure needs a show cue. Case 001 shipped Nickel created
hiddenwith noshow, so his lines played to an empty frame. Before render, check each hiddennickel/lead figure has ashowor arisecue in a beat where it speaks. - Short cut in the same generator (see § The 30 s short).
- Open: the top-left case tag. Case 001 shows
THE CATCH · CASE 001(CASE_TAG, screen layer, every scene). The owner liked everything except this and wants it rethought. Don't copy it into a new case unchanged; propose options (see handover § Open) and use the owner's pick.
First frame and cover (owner 2026-11-01, after YouTube Studio data)
The Case 001 short lost 72% of feed viewers in the first second: it opened on scene-setting ("This is fifty dollars…") instead of the claim.
- Every video starts on its cover, long and short. Frame 0 is the cover: 3–6 words of claim
or question in big type (
phrase,hidden: false,reveal: "none"), the hero object in the cone, Nickel pointing with his beam on it. No empty set, no fade-in, no silent hold: the first line is spoken over the cover (a smallfocuson the hero object), then the story cuts away. No brand anywhere on a cover. - Two covers per case, one per format, each built once in the generator and used twice:
cover_long()/cover_short()return the figures; the opening scene uses them as its figures (long: a one-line scenecoverbefore the cold open; short: its first scene), and a silent one-scene skit renders the still from the same list, so cover and frame 0 never drift apart:<id>-cover(16:9) →out/<id>-cover.png: upload as the long video's thumbnail.<id>-short-cover(9:16) →out/<id>-short-cover.png: the Short's thumbnail. Shorts pick it from the video's frames (pick frame 0) or take this file as a custom upload.- Render:
pnpm still Skit out/<id>-cover.png --frame=40 --props='{"skit":"<id>-cover"}'(same for-short-cover). Covers need no voice.
- Long cover = curiosity question ("Who had my $50 for 4 years?"); short cover = the claim ("Your gift card is a loan to the store"), which is also the short's first spoken line (under 3 s). The human story comes after the claim in the short, or only in the long cut.
- Check frame 0 of both cuts and both cover PNGs before any render is handed back.
The 30 s short (every case)
Every 16:9 case also ships a ≤ 30 s 9:16 cut for the Thin Lines Theory channel that sends
viewers to the long video: the hook, one clue with its best number, Nickel's reaction, then
"Full case, linked below." Never give away the catch. Same sets and kinds (the night sets fit
9:16). meta.links first item { "label": "▶ Full case (<title>)", "url": "{{LONG_URL}}" };
after upload also set the Short's YouTube "related video" to the long one and pin a comment with
the link. Worked example: public/skits/the-drawer-short/.
storyboard.md format
One row per beat: #, time, speaker, line, picture (set, objects, Nickel/lead action), camera,
cue word → action, SFX, music. Then a "Build needs" list (sets, kinds, props, poses) with
planned/exists.
Checks before render
pnpm check <id>: variety (kinds per length, runs, 4–8 s changes, 3 s hook), numbers match
datasets, wording. Frame sheet as in the explainer skill; reject any frame that is a chart on a
flat fill, a frame with two Ember elements, or 8 s without change. Check the frames of every
scene that names a company for anything that looks like its branding, and add the result to the
case's Brand check line in sources.md.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

