youtube-scriptwriter

YouTube Scriptwriter

Write scripts for Ray's YouTube channel (@RAmjad) — Claude Code tutorials and AI tooling content aimed at developers and increasingly non-developers.

Before writing anything

Read the reference scripts in references/ — these are Ray's proven hits. Study their structure before drafting:
- auto-dream-script.md — 90k views in 2 days (gold standard)
- skills-2.0-script.md — 40k views in 2 days
- btw-and-fork-script.md — 30k views in 1 day
Read socials/youtube/improvements.md in the project root — active production improvements. Apply any that are relevant.
Read projects/agentic-coding-school/promotion-playbook.md — the 7-move promotion playbook derived from the 2026-05-23 transcript-pattern analysis (demo surface, objection-handler script, named free artifact, micro-scarcity, single CTA, free-video cross-link, refund stat). The positioning question batch below pulls from this file. If socials/youtube/free-class-videos.md exists, also read it — it's the index of free class videos used to pre-pick the cross-link in rule #6.
If the video topic relates to an existing script in socials/youtube/videos/, read it for context on what's already been covered.

Before drafting — ask the user 4 upfront questions

These four decisions cascade through the whole script and are cheap to answer upfront but expensive to retrofit. Ask all four in a single AskUserQuestion batch before writing any prose. Do not guess — guessing wrong means rewriting the script. If the answers are already unambiguous from context (e.g., the user already said "let's do a pillar" or "newsletter push only"), skip that question, but default to asking.

Q1 — Format class. Which template should this video pull from?

Single-feature deep dive (8-12 min, ~45K view ceiling) — the default for changelogs and feature announcements. Use the /btw, Skills 2.0, or Auto Dream scripts as references.
Thesis / leak / industry take (caps ~17K views) — flag this ceiling upfront so the user calibrates expectations. These videos optimize for thesis-vs-thesis comparison, not raw view count.
Pillar (30-45 min lifestyle storytelling, ~20K views but 3.16% masterclass CTR, 5x per-view revenue) — no single-feature anchor; compress multiple techniques into a "how I actually use Claude Code" narrative. Top 0.01% is the template. Aim for ~1 in 5 videos in this format — they're the highest-converting on the channel.

Q2 — Pitch goal. Which monetization system is this video serving?

Masterclass — soft anchor at 1:30 + closing urgency pitch. Requires a real deadline and price change.
Newsletter — single clean close at the end. No masterclass mention anywhere in the script.
None — pure growth video, channel trailer, or collab.

Never stack newsletter + masterclass in one video. March data is unambiguous: Auto Dream split both weakly and made $6K on 97K views; Kills OpenClaw committed to masterclass only and made $13.7K on 30K views. Stacking was −46% vs single-system pitching. If the user says "both," push back and ask which audience this specific video attracts.

Q3 — Coined term candidate. What phrase is this video trying to put into the viewer's vocabulary? The top 3 videos by views all coined a term ("context pollution," "capability uplift vs encoded preference," "sleep-deprived agent"). The next 7 top performers did not. If the user can't name a candidate in 10 seconds, pause and brainstorm before drafting — this is the single biggest breakout lever and the hardest to retrofit.

Q4 — Live deadline + price escalation. Only ask if Q2 is masterclass. What's the concrete deadline and price change to cite in the closing urgency pitch? The Kills OpenClaw template requires real numbers: "lifetime plan retires [date], price goes up to [amount] after." If there's no real deadline, fall back to a softer close — don't manufacture fake urgency, the audience can smell it.

After Q1-Q4 are answered, fire a second AskUserQuestion batch (positioning batch) drawn from projects/agentic-coding-school/promotion-playbook.md. These three decisions cascade through the script's demo and close sections; they're cheap to lock in upfront and expensive to retrofit:

Q5 — Demo surface. Which project is on screen during the main demo? Default is the masterclass landing page or its MCP server (rule #1 in the playbook). Only swap if the topic structurally demands a different surface (HyperWhisper for a Mac feature, RayOS for a daily-workflow video). The top-3 RPV videos all used the masterclass as the demo substrate. Pitch and demo collapse into the same section when this is done right.

Q6 — Free class video cross-link. Which free video in the class pairs with this topic? This goes in the close (playbook rule #6). Pre-pick from socials/youtube/free-class-videos.md if it exists; if not, ask Ray to name one or flag that a free companion video should be recorded. The verbatim close shape is in the playbook. If Ray can't name a pairing, fall back to the MCP install ("install the masterclass MCP, it'll recommend videos based on your workflow") — still a class-platform tripwire.

Q7 — Named free artifact. Ask if Q2 is masterclass or newsletter. What's the named free artifact tied to this video — PDF cheat sheet, downloadable skill, checklist, dashboard template? Generic "free videos in my newsletter" is wallpaper (playbook rule #3). Must be specific. If Ray can't name one in 10 seconds, pause and brainstorm — for newsletter-goal videos this is the load-bearing pitch element.

Only after Q1-Q4 AND Q5-Q7 are in hand should you start drafting. The playbook's rules #2 (verbatim objection-handler), #4 (micro-scarcity), #5 (single CTA enforcement), and #7 (refund stat in close) are applied automatically during drafting — no questions needed; just bake them in.

The Meta-Pattern (what all 3 hits share)

These three scripts look different on the surface — one's about memory, one's about testing, one's about context management. But they share a deep structure that explains why they all broke out. This is the formula.

1. Name the Invisible Problem

All three scripts start by naming something the viewer feels but can't articulate:

Auto-dream: "Your memory files are full of stale, contradictory notes" (everyone noticed this but nobody named it)
btw/fork: "Context pollution" — coined a term for attention degradation mid-session
Skills 2.0: "Building a skill was pure vibes" — named the lack of testing

The pattern: find the pain that has no name yet, and name it. This is why people share the video — they finally have vocabulary for their frustration.

Coin a term. The data here is stark: the only 3 videos that coined a memorable term ("context pollution", "capability uplift vs encoded preference", "sleep-deprived agent") are the top 3 by views. None of the next 7 top performers coined anything. Every script should try to create at least one phrase viewers will screenshot and reuse. "Context pollution" is the kind of term that spreads beyond the video.

2. Human Analogy as Structural Backbone (not decoration)

Every hit maps the technical concept to a human experience, and the analogy isn't a throwaway line — it's the organizing principle of the entire video:

Auto-dream: Memory consolidation = REM sleep. Every section references sleep/dreaming.
btw/fork: Context window = your desk. /btw = asking without putting paper on it. Fork = second desk. Rewind = clearing the mess.
Skills 2.0: Evals = teacher grading an exam. Capability vs preference = two types of knowledge.

Ask yourself: "What does this feature do that humans also do?" Build the video around that mapping. If the analogy only appears once, it's not doing enough work.

3. 3-Beat Hook (first 20-30 seconds)

Use discovery framing, not announcement framing. March 2026 confirmed this pattern hard. Every March video that opened with personal discovery beat its peers:

Auto Dream (97K views): "I was just doing some vibe coding and I noticed something weird"
/btw (39K views): "Claude Code added a feature I didn't even realize I needed"
Kills OpenClaw (30K, $13,776 — best revenue of the month): "I woke up a few minutes ago and checked the changelog"

Mid-tier March videos that opened with "Anthropic Just Dropped X" verbally (even when the title used that phrase) averaged 30–60% fewer views than the discovery-framed ones. The title can say "Anthropic Just Dropped X" — the spoken opening cannot. Discovery framing ("I stumbled onto this") makes viewers lean in. Announcement framing ("here's the news") makes them evaluate whether to stay.

Every script opens with three beats:

Discovery or novelty signal — frame it as something Ray found, not something Anthropic announced
What it actually is — one concrete sentence explaining the feature
The analogy/reframe that makes it click — connect to the human experience or name the problem

4. The Problem Before the Solution (30s - 2min)

Before explaining how the feature works, make the viewer feel the pain it solves. Use second person — "you've probably run into this." Be brutally specific about the failure mode, not abstract:

Auto-dream: stale memory files, contradictory notes, relative dates that decay
btw/fork: noise in conversation, off-topic messages degrading all future outputs
Skills 2.0: "try it a few times, go 'yeah that seems to work,' and move on"

Concrete details build trust. Vague problems feel like filler.

Use the Problem → Old Workaround → New Solution triple. The top performers don't just show the problem and the solution — they add a middle step showing Ray already tried to solve it the hard way. This builds credibility and makes the new solution feel like relief:

btw/fork: Problem (interrupting pollutes context) → Old workaround (--fork-session, "clunky") → New solution (/btw)
Remote control (36k): Problem (can't use Claude from phone) → Old workaround (tmux + Tailscale + Terminus, "a hassle") → New solution (claude rc)
Kills OpenClaw (31k): Problem (remote was passive) → Old workaround (OpenClaw) → New solution (/loop cron)

Three acts inside the problem section. The "old workaround" step is what separates 30k+ videos from 15k ones — it proves Ray lives in the problem, not just reports on it.

5. Single Feature Deep Dive (2min - 6-7min)

One feature, explained thoroughly. Show the actual internals — real prompts, real code, real terminal output, real benchmark tables. All three hits do this:

Auto-dream: the literal system prompt, the three phases, the trigger conditions
btw/fork: the actual CLI flags, the tradeoff matrix
Skills 2.0: real benchmark tables with pass rates and token counts

Structure in numbered phases or clear comparison sections. People need scaffolding to follow technical content.

Concept-first, not demo-first. The 40k+ videos explain the feature through analogies and show real internals (prompts, tables) but don't do live building. The 25-35k tier videos (remote control, kills OpenClaw, subagents) have extended live-coding sequences — setting up Telegram bots, SSHing into servers. Live demos are fine but they cap out around 30k. The videos that break 40k lead with the idea, not the tutorial. If there's a demo, narrate over it rather than building live.

6. The "Ray Thinks Deeper" Move

All three scripts have a moment where Ray goes beyond the announcement and adds original analysis:

Auto-dream: connects to Mastra's Observational Memory (broader industry trend)
btw/fork: "btw is the inverse of a subagent" (original insight not in the docs)
Skills 2.0: "Anthropic kind of buries the lede" + the insurance claim example showing A/B is the wrong question for procedural skills

This is the section that makes people subscribe. It shows Ray isn't just parroting release notes — he's thinking about what the feature means. Every script needs one of: a broader trend connection, an original insight not in the docs, or a critique/gap the announcement missed.

7. Practical Takeaway (brief)

Brief, actionable. Where to enable it, what to configure, what it means for their workflow. Keep it short — the value was already delivered.

8. The Pitch Layer (separate from content structure)

Content structure is optimized for views. The pitch layer is optimized for revenue. They are different systems and must be planned separately. March 2026 data settled this:

Pitch tactic	With	Without	Delta
Masterclass CTA in first 90 seconds	$11,687 avg	$4,692 avg	+149%
Urgency + price escalation language	$13,776 (Mar 7)	$5,880 avg	+134%
"Best at their company" social proof	$12,142 (Skills 2.0)	$6,064 avg	+100%
Newsletter-first closing	$4,601 avg	$8,489 avg	−46%

The old "single CTA at end, no mid-roll CTAs" rule was costing revenue. Auto Dream followed that rule on 97K views and made only $6,029. Kills OpenClaw broke it (mid-video soft anchor + closing urgency pitch) and made $13,776 on 30K views.

Apply the Mar 7 template to every regular video:

Soft anchor at 1:30–2:00 (after the hook lands, before the deep dive):

"Before we continue — this video is sponsored by myself and my Claude Code Masterclass. Over 1,500 engineers from companies like [X] have taken it and many are now the best Claude Code user at their company. The lifetime plan retires [date] and the price is going up after that. Link below."

Closing urgency pitch (tied to the video topic):

"So [feature summary]. If you want to go deeper on this, the Masterclass covers [specific deliverable]. The lifetime plan is gone after [date] and the price goes up. Link below."

Rules:

Never newsletter-first on videos with a live masterclass deadline (March penalty: −46%)
Always include a concrete deadline + price escalation, not a vague "check it out"
Tie the closing pitch to a specific deliverable the video hinted at
One pitch system at a time. If newsletter is the goal, skip the masterclass pitch entirely rather than stacking both weakly (Auto Dream's mistake)
If a video crosses 30K views in 48 hours and the pitch was weak, add a pinned comment with urgency + masterclass link. Auto Dream would have captured ~$14K more with this single action.

Adapting the Formula by Scope

The three hits show the formula scales to different video lengths:

Video	48hr Views	D1 Subs	D1 Avg Watch	Retention @50%	Has Broader Trend?
Auto-dream	86,662	1,113 (1.90%)	2:52	36%	Yes (Mastra)
Skills 2.0	37,635	343 (1.21%)	4:06	30%	Yes (eval=skill future)
btw/fork	34,495	383 (1.45%)	2:50	33%	No

8-10 min (tight): Skip the full demo. Use the broader trend OR the original insight, not both. btw/fork is the template.
10-12 min (standard): Include the broader trend section AND either a demo or a deeper original analysis. Auto-dream is the template.
12+ min (deep): Full demo walkthrough + broader trend + original critique. Skills 2.0 is the template.
30-45 min (pillar): Lifestyle storytelling opening, no single-feature anchor, compress multiple techniques into one "how I actually use Claude Code" narrative. Mar 2 Top 0.01% (41:31) is the template — it had a 3.16% site click-through rate, 4–5x any single-feature video in March, and produced $10K revenue on only 20K views. Open with memorable lifestyle stories (Claude Code reverse-engineered a dating app API, script to auto-reconnect to Starbucks WiFi). Use for ~1 in 5 videos — it's the highest-converting format the channel has.

Pacing Guidelines (from improvements.md)

Speak ~50% slower than typical AI tutorial pace
Let visuals linger 2-3 seconds after referencing them
Keep natural pauses — don't script out every breath
Add clear section titles/breaks between segments
Use progressive reveal over static slides where possible

What NOT to Do

No feature roundups. Focused single-feature videos averaged 45K views in March vs 17K for thesis/leak/pillar formats — a 2.6x gap. If the announcement has 5 features, pick the most interesting one.
No "Anthropic Just Dropped X" spoken openings. The title can use that framing; the first spoken line cannot. Every March video with an announcement-style verbal opening underperformed the discovery-framed peers on the same week by 30–60%.
No abstract intros. Don't open with "In today's video we're going to..." — open with the discovery beat.
No repeated title templates within 60 days. "Nobody Knew They Needed" was used twice in March (/btw Mar 11 at 39K → Auto Dream Mar 24) and once in Feb. Reuse compresses the novelty signal. Track recent titles before proposing a new one.
No newsletter-first closings on videos with a live masterclass deadline. March penalty: −46% revenue. Newsletter-first averaged $4,601 per video; masterclass-first averaged $8,489.
No stacking 3 uploads in 3 days. Mar 19/20/21 combined drove $8,500 across 5 unique days — less than the single Mar 11 /btw upload ($8,788 on its own). Overlapping 3-day windows cannibalize each other.
No slides-first approach. Write the script first, plan visuals second. The verbal explanation carries the video; visuals support it.

Output Format

Write the script as a markdown document with:

Timestamp ranges for each section (estimated)
Section headers matching the formula above
Actual spoken words (not bullet points — write what Ray will say)
*Italicized stage directions* for screen content and visuals
Save to socials/youtube/videos/<kebab-case-title>.md with frontmatter:

---
tags: [youtube, script, claude-code]
status: draft
date: YYYY-MM-DD
---

Title Generation

Generate 3 title options using different formulas:

Bold claim + anthropomorphism/personification
Bold claim + specificity
Curiosity gap + exclusivity

Present them in a table at the top of the script, like the auto-dream reference does.