content

Content Studio — turning modules into watchable content

This directory turns curriculum modules into published content: long-form YouTube lessons, vertical Shorts/Reels, and live-coding streams. Everything here is an editable source — a storyboard, a script, a run sheet — never a locked export. You (or an editor) record from these.

Brand rules come from ../docs/BRAND.md and the parallel-programming-brand skill. Colors, lower-thirds, and code styling are fixed so every video looks like the same series.

---

Why this exists

A module README teaches a reader who chose to sit down and study. Video has to earn attention in three seconds and hold it without the viewer being able to re-read a line. That is a different craft, so each module gets a separate content pack rather than someone reading the README aloud.

The pipeline:

%%{init: {
  'theme': 'base',
  'themeVariables': {
    'primaryColor': '#F8FAFC',
    'primaryTextColor': '#0F172A',
    'primaryBorderColor': '#0891B2',
    'lineColor': '#64748B',
    'secondaryColor': '#E0F2FE',
    'tertiaryColor': '#EEF2FF',
    'fontFamily': 'Inter, Segoe UI, sans-serif'
  }
}}%%
flowchart LR
  module["Module README<br/>9 sections"] --> extract["Extract the ONE idea<br/>+ the measured number"]
  extract --> long["Long-form script<br/>8-12 min"]
  extract --> shorts["Shorts / Reels<br/>3 x 30-60 s"]
  extract --> live["Live-coding run sheet"]
  long --> record["Record: screen + terminal + diagram"]
  shorts --> record
  live --> record
  record --> publish["Publish + link back to the module"]

  classDef trackA fill:#0891B2,stroke:#0F172A,color:#fff
  classDef neutral fill:#F8FAFC,stroke:#0891B2,color:#0F172A
  classDef trackC fill:#F59E0B,stroke:#0F172A,color:#0F172A
  class module,extract neutral
  class long,shorts,live trackA
  class record,publish trackC

Loading

Every content pack starts from the module's measured result, because a real number on screen is what makes a GPU video credible instead of generic.

---

Directory layout

content/
├── README.md                  <- this guide
├── STORYBOARD_TEMPLATE.md     <- copy this per module
├── BRAND_OVERLAY.md           <- lower-thirds, title cards, code theme, captions
├── medium/PLAN.md             <- Medium publication starter plan (long-form text channel)
├── social/                    <- LinkedIn / Shorts / Medium packs + HISTORY chain
│   ├── HISTORY.md             <- recurrent post tracker (update every pack)
│   └── NNN-<slug>/            <- feed-length cut of a measured idea
└── <module-id>-<slug>/
    ├── long-form.md           <- 8-12 min YouTube script (shot-by-shot)
    └── shorts.md              <- 3 vertical Shorts/Reels scripts

Social packs are authored with the linkedin-shorts-maker skill. See social/README.md and always update social/HISTORY.md. Long-form text publishing follows medium/PLAN.md.

---

The formats

Format	Length	Aspect	Goal	Success signal
Short / Reel	30–60 s	9:16	One idea, one number, one "huh!"	Watched to the end, saved
Long-form	8–12 min	16:9	Teach the module's core skill	Retention past the demo
Live coding	45–90 min	16:9	Show the messy real process	Chat questions answered

Shorts: the 5-beat structure

Shorts fail in the first second or not at all. Use this beat sheet:

1. Hook (0–3 s) — a claim or a number that sounds wrong. "This one-line change made the GPU 8× faster."
2. Stakes (3–8 s) — why the viewer should care. "Same math. Same kernel. Only the memory access changed."
3. Reveal (8–35 s) — the actual idea, on screen, with the code diff or diagram.
4. Proof (35–50 s) — the measured number in the terminal. Never a claim without the run.
5. CTA (50–60 s) — "Full module and runnable code in the description."

Rules: no intro card (it kills retention), burned-in captions always, one idea per Short, and the number must come from a real run.

Long-form: the arc

Follow the module's own logic — it is already pedagogically ordered:

Minutes	Beat	Sourced from module section
0:00–0:30	Cold-open the payoff (show the final speedup first)	Section 5
0:30–1:30	The analogy — make a layman get it	Section 1
1:30–3:30	First principles: derive why the naive way breaks	Section 2
3:30–6:30	Live code the fix, dual-vendor if relevant	Sections 3–4
6:30–8:30	Run it, read the profiler, tie to the roofline	Section 5
8:30–10:00	Honest tradeoffs and when NOT to do this	Section 6
10:00–11:00	Where this shows up in real ML systems	Section 7
11:00–12:00	The drill for the viewer + what's next	Section 9

Cold-opening with the result is deliberate: viewers stay for a payoff they have already seen is real, and the analogy lands better once they know what it is explaining.

---

Recording setup

• Terminal: dark, background #0F172A, JetBrains Mono, ≥16pt. Viewers on phones cannot read a 12pt terminal — this is the single most common mistake in GPU content.
• Editor: same font; hide the file tree while zoomed on code.
• Diagrams: render the module's Mermaid diagrams (GitHub and the MkDocs site both render them) and screen-capture. For a curriculum-overview shot, open the interactive curriculum dashboard canvas in Cursor (see BRAND_OVERLAY.md for which shots work on camera).
• Infographics: the assets/ SVGs are already animated and already on brand, so open one in a browser tab and capture it rather than rebuilding it in an editor. Name the asset id in the shot row. If a file has to be handed over, derive it with bash scripts/export_visual.sh <asset> — never commit the render.
• Interactive shots: the concept players put a control on screen, which makes a much better reveal than a fixed drawing when the lesson is "watch this number move". live/concepts/coalescing.html is the one that exists today.
• Live GPU: always show rocm-smi / nvidia-smi once, so viewers know it is real hardware.
• Audio matters more than video quality. A cheap mic close to your mouth beats a good mic far away.

---

Publishing checklist

• Title names the concrete outcome, not the topic ("Why your GPU kernel is 8× slower than it should be", not "Memory coalescing").
• Thumbnail: one number + one 3-word phrase, brand cyan #0891B2 on slate #0F172A.
• Description links the module folder and the repo.
• Burned-in captions on Shorts; SRT on long-form.
• Chapters on long-form (use the arc table above).
• The measured number on screen matches what the code actually prints.
• Any hardware claim states the GPU it ran on.

---

Content packs

Module	Pack	Hook number	Source of number
A01 Foundations	A01-foundations/	Kernel is 44 µs of a 72 ms run (0.06%)	rocprofv3 profile run
A02 Memory hierarchy & coalescing	A02-memory-coalescing/	700 → 3,400 GB/s (4.8×)	A02 Section 5 reference run

Each pack has a long-form.md (shot-by-shot YouTube script with chapters and a description block) and a shorts.md (three vertical scripts). Build the next pack by copying STORYBOARD_TEMPLATE.md.

All 30 curriculum modules are authored, but only these two have content packs. That is deliberate: a pack needs a measured number to lead with, and inventing one would break the rule the whole curriculum rests on. The next four are prepared to the point where the only missing piece is the run:

Module	Blocked on	Lead visual	Beat it belongs in
A04 Parallel reduction	Stage 1 to stage 9 spread from the reduction lab	asset: reduction-tree	Reveal — the depth win, then the idle cost
A06 Tiled matmul	naive vs tiled vs register-blocked GFLOP/s from the matmul lab	asset: tile-reuse	Reveal — the two traffic bars, to scale
A07 Softmax & fusion	the poisoned row printing nan, then not	— (terminal is the visual)	Hook — the failure on screen
B02 Roofline	the knee the intensity sweep finds on your own part	asset: roofline	Proof — your kernels plotted on your ceilings

Each lead visual already exists in ../assets/, so the pack is a script plus a number, not a script plus a number plus artwork. Capture a visual by opening its SVG in a browser tab, or derive a file with bash scripts/export_visual.sh <asset> gif 9x16 if an editor needs one.

Re-run the lab on your own GPU before recording and quote your own numbers on screen. The figures above are from reference runs on AMD MI-series hardware; they will differ on yours, and an unattributed number is worse than no number.