A YouTuber pasted his voice corpus into a Claude Project, drafted a script with the same prompt he'd been using for newsletters, and watched the model produce 1,400 words of perfectly-cadenced written prose that died on camera in 90 seconds. The voice corpus was right. The prompt was wrong. Voice is surface-dependent - newsletter cadence is not script cadence is not thread cadence - and a single master system prompt that tries to be all three produces a hybrid that fits none. This lesson builds the three prompts that operationalize your corpus: one for newsletter drafting, one for video script drafting, one for social thread drafting. Each is 800-1,500 words, lives version-controlled in your brand-memory store, and shifts every downstream draft from 15-40% in-voice to 70-80% in-voice on the first generation. Total build time: 2-3 hours one-time.

Why Three System Prompts, Not One

The temptation is to write one master system prompt that handles everything. This fails because voice is surface-dependent. Your newsletter voice has a particular cadence (longer paragraphs, dropped-in personal moments). Your video script voice has different cadence (shorter sentences, spoken-rhythm-driven). Your social thread voice is different still (compressed, hook-driven, line-break-aware). A single system prompt that tries to be all three produces a hybrid that fits none.

The three-prompt approach acknowledges that the voice corpus from Lesson 1.1 contains all three voices, but each system prompt explicitly directs the model to weight toward one. Newsletter prompt says "draft in the newsletter register from the corpus." Script prompt says "draft in the spoken-script register." Thread prompt says "draft in the social-compression register." Same corpus, three surface-specific applications.

The other reason: production-grade system prompts include surface-specific rules. The newsletter prompt bans "let's dive in" openers; the script prompt bans "in today's video we're going to be diving into"; the thread prompt bans "🧵 thread incoming" and single-emoji-per-line patterns. Surface-specific forbidden phrases require surface-specific prompts.

The Production-Grade System Prompt Anatomy

Each of your three prompts follows the same six-part structure:

1. Audience description (1 paragraph) - who reads this surface, what they're tired of, what they value
2. Voice corpus reference (1 sentence) - point to the corpus document; tell the model to weight toward the surface-specific subset
3. Do / Don't rules (10-15 explicit items) - the negative space; cadence preferences; word bans; structural preferences
4. Structural template (3-5 lines) - the canonical structure for this surface (cold open → claim → evidence → counter → resolution → CTA)
5. Forbidden phrases (10-20 items) - surface-specific slop tells; updated quarterly as audience pattern-detection evolves
6. Output format requirement (1-2 lines) - what the model returns (just the draft prose, no preamble, no "Here is your draft" framing)

Total length: 800-1,500 words per prompt. Sits in the system prompt field of the Claude Project / Custom GPT / Gemini Gem. Version-controlled in a text file in your brand-memory store so you can iterate.

Prompt 1: Newsletter Drafting

The canonical L2 newsletter prompt structure. Customize for your audience and voice; the template below is the starting point.

=== NEWSLETTER DRAFTING SYSTEM PROMPT v1 ===

AUDIENCE
You are drafting a newsletter for [audience: e.g., "newsletter operators at 5K-30K subs running paid tiers"]. They are tired of generic AI-cohort content and respond to specific named cases, real dollar amounts, dated events. They read on mobile mostly, often during a Tuesday commute. Reply rate is the trust signal that matters.

VOICE CORPUS REFERENCE
Reference the attached voice corpus. Weight toward the newsletter-tagged pieces. Match cadence variation, hedge-word density (low), specificity density (high).

DO / DON'T RULES
- Never use "let's dive in," "in this article," "without further ado," or "in today's rapidly evolving landscape"
- Never use em-dash parallelism ("It's not just X - it's Y")
- Never close with "what do you think? let me know in the comments"
- Never use "leverage" (as verb), "unleash," "transform," "synergy," "robust"
- Never use 4+ hedge words per page ("perhaps," "arguably," "some might say")
- Always cite specific named sources for any stat
- Always include at least one specific dollar amount or count in the first 200 words
- Always write the cold open as a specific scene or moment, not abstract claim
- Vary sentence length: do not produce 5 consecutive sentences within ±15% of each other
- The second paragraph rhythm must match a real recent newsletter from the corpus
- One opinion held with conviction per issue; the rest is reportage + analysis

STRUCTURAL TEMPLATE
1. Cold open: specific scene / dated moment / named character (1-2 paragraphs)
2. The claim: what this issue argues, in plain language (1 paragraph)
3. The evidence: 2-3 specific named cases, dollar amounts, or studies (3-5 paragraphs)
4. The counter: where the reader might disagree, addressed honestly (1-2 paragraphs)
5. The resolution: what the operator should do this week (1-2 paragraphs)
6. The close: a specific take, not a question to the reader

FORBIDDEN PHRASES (quarterly-updated)
- "let's dive in"
- "without further ado"
- "in today's rapidly evolving landscape"
- "what do you think?"
- "let me unpack that"
- "to put it simply"
- "buckle up"
- "transform"
- "leverage" (as verb)
- "unleash"
- "robust"
- "synergy"
- "It's not just X - it's Y" (em-dash parallelism)

OUTPUT
Return the draft prose only. No preamble. No "Here is your draft." No meta-commentary.

=== END ===

Paste this into your Claude Project / Custom GPT / Gemini Gem system-prompt field. Customize the audience description and a few of the do/don't rules to match your actual voice and audience. Don't strip the structure - every section is doing operational work.

Prompt 2: Video Script Drafting

The L2 YouTube script prompt is structurally similar but tuned for spoken cadence and the 2026 algorithm gates (3-second hold + ~70% watch-through).

=== VIDEO SCRIPT DRAFTING SYSTEM PROMPT v1 ===

AUDIENCE
You are drafting a script for [creator]'s YouTube channel targeting [audience: e.g., "operators at the 5K-50K sub range running long-form content businesses"]. The viewer is watching on a phone in 75% of cases; the hook is the 3-second hold gate; the overall watch-through target is 70%+ for the algorithm-boost layer. Captions are doing 85% of the load-bearing work because the watch is silent.

VOICE CORPUS REFERENCE
Reference the attached voice corpus. Weight toward the script-tagged and podcast-monologue-tagged pieces (spoken-rhythm, shorter sentences). Match the spoken-cadence subset, not the written-newsletter subset.

DO / DON'T RULES
- Never use "in today's video we're going to be diving into"
- Never use "let me know in the comments below" close
- Never use "as always, like and subscribe" without a specific reason
- Never use "in this video I'll cover [list]" preamble
- Write for spoken delivery: sentences under 18 words median, with 2-3 longer sentences for rhythm
- Cold open in the first 3 seconds with a specific scene or claim, not a brand intro
- The 15-second mark should land a payoff that justifies the next 12 minutes
- Use the operator's actual stories and named examples, not generic illustrations
- Pacing: payoff every 60-90 seconds; if a paragraph doesn't pay off in 60 seconds of spoken time, cut or restructure

STRUCTURAL TEMPLATE
1. 0-3s: Specific hook (scene/claim/number, no preamble)
2. 3-15s: Payoff that justifies the rest (15-second retention checkpoint)
3. 0:15-2:00: The setup (what is this about, why now)
4. 2:00-10:00: The body (3-5 specific cases or steps, with named examples)
5. 10:00-12:00: The resolution + CTA tied to specific next action

FORBIDDEN PHRASES (quarterly-updated)
- "in today's video"
- "hey everyone, welcome back"
- "before we get started"
- "let me know in the comments"
- "as always, like and subscribe"
- "smash that subscribe button"
- "without further ado"
- "let's get into it"
- "diving into"
- "buckle up"

OUTPUT
Return the script prose only. Mark spoken cues in brackets [pause], [cut to chart], [emphasize] where useful. No preamble, no meta-commentary, no "Here is your script."

=== END ===

Prompt 3: Social Thread Drafting

The thread prompt handles X/Twitter, LinkedIn, and Threads variants. Compressed format, hook-driven, surface-aware.

=== SOCIAL THREAD DRAFTING SYSTEM PROMPT v1 ===

AUDIENCE
You are drafting threads for [creator]'s social distribution, targeting [audience]. Platform-specific weighting: X audience is harshest detector of AI slop (Lesson 2.3 of L1); LinkedIn rewards concrete named outcomes; Threads rewards held conviction; Bluesky rewards prose form.

VOICE CORPUS REFERENCE
Reference the attached voice corpus. Weight toward the thread-tagged and short-form-tagged pieces. Match the social-compression register.

DO / DON'T RULES
- Never use single-emoji-per-line thread pattern (💡 first 🎯 second 🚀 third)
- Never use "thread incoming 🧵" preamble
- Never use "Here's the truth nobody is telling you" or "What 99% of [audience] get wrong"
- Never use generic numbered lists with no specificity
- Never close with "RT if you found this useful"
- For X threads: each tweet must work as standalone read; first tweet is the hook, must include a specific number or named case
- For LinkedIn: 3-7 paragraphs typical; named outcome up front; arrow-bullet pattern (→) sparingly, not uniformly
- For Threads: lead with held conviction; no hedging in the follow-up line
- For Bluesky: prose form; no list structure; deliberate take

STRUCTURAL TEMPLATE (X thread, default)
1. Tweet 1: Hook (specific number / named case / dated event)
2. Tweets 2-4: The setup (what is this about, why now)
3. Tweets 5-8: The body (3-5 specific cases with named details)
4. Tweet 9-10: The resolution + take

FORBIDDEN PHRASES (quarterly-updated)
- "Here's the truth nobody is telling you"
- "What 99% of [audience] get wrong"
- "Hot take:" (without an actual take following)
- "POV:" (when not POV content)
- "Is it just me or..."
- "Controversial opinion 🔥"
- "Buckle up"
- "Thread incoming"
- "RT if you found this useful"

PLATFORM ADAPTATIONS
- If output type = X thread: 8-10 tweets, each under 280 chars
- If output type = LinkedIn: 3-7 paragraphs, longer-form, named outcome up front
- If output type = Threads: 1-3 posts, direct emotional registration
- If output type = Bluesky: 1-2 prose posts, no list format

OUTPUT
Return the thread/post text only. Mark tweet breaks with "===" for X threads. No preamble, no meta-commentary.

=== END ===

The Three Prompts Side-by-Side

Decision rule: Use the newsletter prompt when the read mode is sit-down with attention budget. Use the script prompt when the consumption is spoken delivery on a phone with captions doing 85% of the load. Use the thread prompt when the reader is scrolling at speed and the hook has to hold in 280 characters.

Composite Case: The Three-Prompt Overhaul

Composite Case: David Okonkwo, B2B Newsletter + YouTube Operator (composite of four operators). David ran one master prompt for 14 months across his newsletter (9,200 subs), YouTube channel (18K subs), and X account. His Tuesday reply rate held at 2.1%; his YouTube AVD was 38%; his X engagement was flatlined at 0.3%. He spent a Sunday in February 2026 splitting the master prompt into three surface-specific prompts - newsletter (1,180 words), script (1,340 words), thread (980 words) - each with its own forbidden-phrase list and structural template. Inside 8 weeks: Tuesday reply rate rose to 3.6%, YouTube AVD climbed to 52%, X engagement hit 1.2%. The single change with the biggest delta: the script prompt's "under 18 words median + 15-second checkpoint" rule pushed AVD up 9 points alone in the first two videos.

How the Three Prompts Work Together

The three prompts share infrastructure (same voice corpus, same brand-memory store, same quarterly review cadence) but differ in operational details. A typical L2 operator workflow:

1. Tuesday newsletter draft - invoke Newsletter prompt with topic + key claim. Output: 80% in-voice first draft.
2. Wednesday video script - invoke Script prompt with topic + research brief. Output: spoken-cadence script ready for record.
3. Thursday repurposing - invoke Thread prompt three times (X variant, LinkedIn variant, Threads variant) from the same core idea. Output: three platform-edited variants in 15 minutes total.

Same corpus, three prompts, three surfaces, one shipping week. This is the L2 productivity unlock that makes the 2-3x cadence outcome operational.

Version Control and Iteration

The three prompts are v1 documents. Save them in your brand-memory store as text files (newsletter-prompt-v1.txt, script-prompt-v1.txt, thread-prompt-v1.txt). When you iterate, increment versions: v1.1 (small rules tweak), v1.2 (added 3 forbidden phrases), v2.0 (structural change).

Iterate based on three signals:

• Output quality. If drafts feel generic, tighten do/don't rules + add forbidden phrases.
• Surface evolution. If your newsletter audience shifts, update audience description + structural template.
• Platform algorithm changes. When YouTube changes its retention dynamics, your script prompt's "15-second checkpoint" line might need adjustment.

Quarterly review aligned with voice corpus re-curation: re-read each prompt, update forbidden-phrases list to current slop tells, adjust audience descriptions if shifted. Version-bump as you go.

The L2 Deliverable

The output of this lesson: three production-grade system prompts (newsletter, script, thread), each 800-1,500 words, loaded into your Claude Project / Custom GPT / Gemini Gem, version-controlled in your brand-memory store. Total build time: 2-3 hours one-time. Iterated quarterly.

Combined with the voice corpus (Lesson 1.1) and the rewrite-loop discipline (Lesson 1.3, next), these three prompts are the operational triple that makes every L2 chapter ship faster while preserving voice.

The System Prompt Economics (Q1 2026)

Per-prompt build time: 2-4 hours one-time per prompt × 3 prompts = 6-12 hours total. Annual refresh: 1 hour/quarter per prompt × 3 prompts = 12 hours/year. Tool cost: Claude Projects $20/mo (already in L2 stack). No additional spend.

Time-recovered per use: 15-30 min per piece from voice-pass overhead reduction. Operator shipping 4 newsletters + 4 scripts + 4 thread-batches/month × 12 = 144 uses/year × 22 min avg saved = 53 hours/year recovered. At $200-300/hr opportunity: $10,600-$15,900/year time-equivalent return on 6-12 hour build.

Failure Modes Specific to System Prompts

One-prompt-fits-all. Operator uses single prompt for newsletter/script/thread. Output drift across formats. Fix: 3 distinct prompts per Lesson 2.1.2 - each calibrated to format conventions.

No voice anchors in prompt. Prompt instructs structure but not voice. Output reads structurally clean but voice-generic. Fix: explicit voice anchors in prompt (3-5 specific operator-voice patterns to enforce + 5-8 patterns to avoid).

Stale prompt. Operator built prompts 12-18 months ago; voice has evolved but prompt hasn't. Fix: quarterly 1-hour refresh per prompt.

Skip the rules layer. Operator includes positive voice anchors but no negative rules ("never use em-dash parallelism", "avoid 'dive in' opener", "no closing question that doesn't invite reply"). Fix: 5-8 negative rules per prompt to block slop tells.

Single-format prompt overload. Operator stuffs newsletter + script + thread instructions into one prompt. Token cost rises; output focus drops. Fix: 3 separate prompts; switch context per task.

The 2026 Industry Context Behind This Lesson

The three system prompts are the operational layer that turns the voice corpus into shipped output across the three primary content formats - and the 2026 model landscape made them dramatically more effective than 2024 equivalents. Claude Projects, ChatGPT Projects, and Gemini Gems all support system-prompt persistence with full corpus context by 2026, which means an 800-1,500 word system prompt + 8,000-15,000 token voice corpus runs as the always-on context for every generation in that Project. Pre-2026, the operator pasted the prompt fresh per conversation; post-2026 it persists, which is how every Tuesday issue and every video script starts at 70-80% in-voice instead of 15-40%. The 2-3 hour build investment amortizes across roughly 200+ generations per year for an active operator.

Three 2026 platform mechanics matter directly for these prompts. The Beehiiv MCP integration (March 2026) reads the newsletter system prompt through the model context, which is how Lesson 2.2.1's 90-minute Tuesday workflow holds voice while incorporating real-time subscriber data. The Castmagic ($120K MRR Q1 2026) and Tella ($500K MRR Q1 2026) production tools per founder-transparency disclosures both accept system-prompt overlays in their 2026 product updates - meaning podcast-derived assets and recorded video segments inherit the operator's voice through the same prompt infrastructure. The FTC May 2026 update to 16 CFR Part 255 made each AI-augmented piece individually accountable for substantiation; the system prompts in this lesson include verification-protocol references that route claim-bearing drafts through Lesson 1.2.4's check.

Why this lesson is the load-bearing infrastructure step for L2: the 50-point voice quality gap between system-prompt-loaded vs. fresh-conversation drafts is exactly the gap between content that compounds (in-voice, retains readers, drives paid conversion) and content that doesn't (off-voice, sounds like everyone else, generates unsubscribes). The 2-3 hour one-time build amortizes across ~200+ generations per year. Operators who skip it operate at a permanent 50-point quality deficit for every downstream piece - and that deficit cannot be recovered by the rewrite loop alone because the rewrite loop's ceiling is bounded by the first-pass register. Skip the prompts, and the rewrite loop tops out at the same 70-80% ceiling fresh conversations produce.

"Positive voice anchors get you 40% of the way; the negative rules get you the other half. 'Never use em-dash parallelism' does more for shipped quality than three more example pieces ever will."

Key Takeaways

• Three production-grade system prompts (newsletter, script, thread), not one master prompt - voice is surface-dependent.
• Each prompt has six sections: audience, voice corpus reference, do/don't rules, structural template, forbidden phrases, output format requirement.
• Length: 800-1,500 words per prompt. Loaded into Claude Project / Custom GPT / Gemini Gem. Version-controlled in brand-memory store.
• Newsletter prompt: specificity-focused, longer paragraphs, dropped-in personal moments. Bans "let's dive in," em-dash parallelism, performative closing questions.
• Script prompt: spoken cadence (under 18 words median), 3-second hook + 15-second checkpoint + 60-90 second payoff cadence. Bans "in today's video," "smash that subscribe."
• Thread prompt: platform-aware (X, LinkedIn, Threads, Bluesky variants); social-compression register; bans single-emoji-per-line, "thread incoming," vague-superiority openers.
• All three reference the same voice corpus but weight toward surface-specific subset.
• Iterate based on output quality, surface evolution, platform algorithm changes. Version-bump (v1.0 → v1.1 → v2.0). Review quarterly with corpus re-curation.
• L2 deliverable: three prompts loaded + version-controlled. Combined with corpus + rewrite-loop discipline = the L2 operational triple.