HRyu
← Library

Projects & Ideas · Field note

AI pipeline

The 20-Agent Script System: How to Build an AI Writing Pipeline That Actually Produces Good Work

A working note — rougher than the essays, kept here for reference.

The 20-Agent Script System: How to Build an AI Writing Pipeline That Actually Produces Good Work

I am creating a course to master Openclaw and Claude AI automations for beginners and also for businesses to get consultation if they don’t have the time to make these automations themselves 

WHAT THIS GUIDE IS

This is a full rebuild of the 20-agent script writing system described in the viral post about generating millions in revenue from AI-written launch scripts. The original post outlined the architecture but did not share the individual agents or their prompts. This guide fills that gap.

You will learn exactly what each agent does, how they connect to each other, and the prompts that power them. By the end, you will have everything you need to set this up yourself in Claude Code or adapt it to whatever AI workflow you use.

Fair warning: this is not a "paste one prompt and get a script" guide. The whole point of the system is that a single prompt produces garbage. This is a pipeline where 20 specialized agents each do one job at a very high standard, and nothing moves forward until it passes a quality gate. That is what makes the output actually good.

HOW THE SYSTEM IS STRUCTURED

The 20 agents are organized into 5 phases. Each phase feeds into the next. Nothing skips ahead.

Phase 1: Research (4 agents) Phase 2: Strategy (3 agents) Phase 3: Writing (6 agents) Phase 4: Quality Control (4 agents) Phase 5: Assembly (3 agents)

In Claude Code, each agent runs in its own context window. The orchestrator (the master script) calls each agent in sequence, passes the output of one into the input of the next, and enforces the rules at every gate.

If you are not using Claude Code, you can run this manually by copying the output of each agent and pasting it as context into the next one. It takes longer but the results are the same.

BEFORE YOU START: WHAT YOU NEED

  1. The brand name and a clear description of what the product does

  2. The product's key features and what makes it different from competitors

  3. The target audience (who is this for, what do they care about, what are they frustrated with)

  4. The format you are writing for (YouTube ad, TikTok, Instagram Reel, product launch video, etc.)

  5. The target length of the final script (30 seconds, 60 seconds, 90 seconds, 2 minutes, etc.)

Put all of this into a single document. This is your Product Brief. Every agent will reference it.

PHASE 1: RESEARCH

The research phase is what separates this system from everything else. Most people skip research entirely and go straight to writing. That is why most scripts sound like they were written by someone who has never used the product, never talked to a customer, and never watched a competing video.

These agents gather real data from real platforms so the script is built on evidence, not guesses.

AGENT 1: YOUTUBE RESEARCH AGENT

What it does: Searches YouTube for the highest-performing videos in your product's category. Identifies the title patterns, hooks, and structures that consistently get views. Finds the ceiling (the best performers) and the floor (where views drop off sharply). Extracts the patterns worth using.

The prompt:

You are a YouTube research analyst. Your job is to identify the highest-performing video content patterns in a specific product category.

Here is the product brief: [PASTE YOUR PRODUCT BRIEF]

Do the following:

  1. Generate 15 keyword searches that a potential customer for this product would type into YouTube. Mix broad terms (the category), specific terms (the problem the product solves), and competitor terms (alternatives they might be considering).

  2. For each keyword, identify what a top-performing video would look like. Think about:

  • What titles consistently appear on videos with the highest views in this space

  • What patterns do those titles follow (numbers, questions, "how to" formats, before/after, controversy, bold claims)

  • What hooks do the top videos open with in the first 5 seconds

  • What video length tends to perform best for this topic

  1. Identify the "ceiling" and "floor" for this category. The ceiling is the view count range of the top performers. The floor is where views drop off sharply. This range tells us what the audience in this space actually watches versus ignores.

  2. List the top 10 title patterns worth stealing. For each one, explain why it works and give 3 examples of how it could be adapted for our product.

  3. List the top 5 hook patterns you see in the best-performing videos. Describe each one in terms of structure: what the first sentence does, what the visual shows, and why it stops someone from scrolling.

Output everything in a structured document with clear sections. This research will be used by the writing agents later, so be specific and include concrete examples, not vague observations.

AGENT 2: REDDIT RESEARCH AGENT

What it does: Mines Reddit for real customer pain, frustration, and desire. Grabs exact quotes from real people talking about the problem your product solves. Finds controversy, strong opinions, and emotional language that can be used in the script.

The prompt:

You are a Reddit research analyst. Your job is to find the raw, unfiltered voice of the customer for a specific product category.

Here is the product brief: [PASTE YOUR PRODUCT BRIEF]

Do the following:

  1. Identify the 10 most relevant subreddits where potential customers for this product would hang out and talk about the problem it solves.

  2. For each subreddit, identify the types of posts and threads that get the most engagement. Look for:

  • Complaint posts (people frustrated with current solutions)

  • Recommendation requests ("what do you use for X?")

  • Comparison threads ("X vs Y, which is better?")

  • Success stories ("I finally found something that works")

  • Controversial opinions (heavily downvoted comments reveal strong feelings)

  1. Extract 20 to 30 exact quotes from real Reddit posts and comments that express:

  • The core pain or frustration the product solves

  • What people wish existed but could not find

  • What they hate about current alternatives

  • The emotional language they use when describing the problem (not corporate language, real human words)

  • Objections or skepticism they would have about a product like this

  1. For each quote, note:

  • The subreddit it came from

  • Whether it was highly upvoted, controversial, or deeply buried

  • Why this quote is useful for scriptwriting (does it reveal a pain point, a desire, an objection, a motivation?)

  1. Summarize the top 5 customer pain points in order of intensity. The most emotionally charged pain point comes first.

Output this as a structured document. The writing agents will use these quotes and pain points as raw material for the script.

AGENT 3: X/TWITTER RESEARCH AGENT

What it does: Pulls high-engagement posts from X/Twitter related to the product category. Identifies the posts where people had strong reactions, fought in the replies, quote-tweeted with opinions. Finds the "nerve" topics that get people emotional.

The prompt:

You are a Twitter/X research analyst. Your job is to find the highest-engagement conversations happening around a specific product category and identify what makes people react.

Here is the product brief: [PASTE YOUR PRODUCT BRIEF]

Do the following:

  1. Generate 15 search queries that would surface relevant conversations on X about the problem this product solves, the product category, and related topics.

  2. For each query area, identify the types of posts that get the highest engagement (likes, retweets, quote tweets, replies). Look for:

  • Hot takes and controversial opinions about the category

  • Posts where the quote-tweet ratio is high (meaning people disagreed strongly enough to add their own commentary)

  • Threads that went viral because they revealed something surprising, contrarian, or deeply relatable

  • Posts from notable figures in the space

  • Customer complaints about competitors that went viral

  1. Extract 15 to 20 high-performing post patterns. For each one, note:

  • The structure of the post (how it opens, what it claims, how it closes)

  • Why it performed well (controversy, relatability, humor, specific numbers, bold claim)

  • The "nerve" it hit (what emotional button did it push)

  1. Identify the 5 most emotionally charged topics in this product space right now. These are the topics where people have the strongest opinions and are most likely to engage with content.

  2. List any specific phrases, slang, or language patterns that are commonly used by the audience on X when talking about this topic. This is the voice of the customer in their most casual, unfiltered form.

Output as a structured document. The writing agents will use this to make the script feel current, emotionally resonant, and worth engaging with.

AGENT 4: RESEARCH SYNTHESIS AGENT

What it does: Takes all the research from Agents 1, 2, and 3 and distills it into a single, organized ammunition file. Identifies the strongest patterns, the most powerful quotes, the highest-impact pain points, and the most promising structural approaches. This becomes the reference document that every writing agent pulls from.

The prompt:

You are a research synthesis specialist. You have received three research documents: one from YouTube analysis, one from Reddit analysis, and one from Twitter/X analysis. Your job is to combine them into a single, actionable reference document for the scriptwriting team.

Here is the product brief: [PASTE YOUR PRODUCT BRIEF]

Here is the YouTube research: [PASTE AGENT 1 OUTPUT]

Here is the Reddit research: [PASTE AGENT 2 OUTPUT]

Here is the X research: [PASTE AGENT 3 OUTPUT]

Create a unified Ammunition File with these sections:

  1. TOP PAIN POINTS (ranked by emotional intensity) List the 5 strongest customer pain points found across all three platforms. For each one, include the best supporting quotes from Reddit and X, and note if YouTube data supports it as a high-performing topic.

  2. LANGUAGE BANK The 20 most powerful phrases, quotes, and expressions found across all platforms that capture how real people talk about this problem. These are not corporate marketing phrases. These are real human words dripping with emotion, frustration, or desire.

  3. PROVEN HOOKS The 10 strongest hook patterns identified from YouTube titles and X posts. For each, write a version adapted specifically for this product.

  4. CONTENT NERVES The 5 topics that generate the strongest emotional reactions. For each, note what makes it controversial or engaging and how it could be incorporated into the script without being manipulative.

  5. COMPETITIVE GAPS Based on the research, what are the top 3 things competitors are NOT saying or doing in their content that represents an opportunity for this script?

  6. STRUCTURAL RECOMMENDATIONS Based on the top-performing content across all platforms, what script structure (length, pacing, format) is most likely to perform well for this product?

This document will be the primary reference for every writing and quality agent that follows. Make it scannable, specific, and full of concrete material they can pull from directly.

PHASE 2: STRATEGY

Before writing begins, these agents make critical decisions about the script's direction, angle, and voice.

AGENT 5: AUDIENCE PROFILER AGENT

What it does: Creates a detailed profile of the person watching the video. Not a generic "target demographic" but a real, specific person with real frustrations, desires, and objections. This profile drives every word choice in the script.

The prompt:

You are an audience strategist. Based on the product brief and research synthesis, your job is to create a vivid, specific profile of the ideal viewer for this script.

Here is the product brief: [PASTE YOUR PRODUCT BRIEF]

Here is the research synthesis (Ammunition File): [PASTE AGENT 4 OUTPUT]

Create an Audience Profile that includes:

  1. THE PERSON

  • Who are they specifically? (not "25-34 males" but "a 29-year-old startup founder who has been trying to get their first product launch to break through for 6 months")

  • What does their day look like?

  • What are they doing right before they see this video? (scrolling X, watching YouTube, browsing Instagram)

  • What mood are they in?

  1. THE PAIN (what keeps them up at night related to this product's category)

  • Their #1 frustration in their own words (pull from the Reddit/X quotes)

  • What they have already tried that did not work

  • What they are afraid of (the fear behind the frustration)

  1. THE DESIRE (what they actually want)

  • The outcome they dream about

  • The version of their life/work/business where this problem is solved

  • The specific moment they would feel the relief

  1. THE OBJECTIONS (what they will think while watching)

  • Their top 3 skepticisms about a product like this

  • What has burned them before

  • The question they will ask before they buy

  1. THE TRIGGER (what would make them act)

  • What proof would overcome their skepticism

  • What emotional beat would tip them from "interesting" to "I need this"

  • What would make them share this video with someone else

This profile should feel like a real person, not a marketing persona. The writing agents will write directly to this person.

AGENT 6: ANGLE SELECTION AGENT

What it does: Reviews all the research and the audience profile and decides on the single strongest angle for the script. The angle is the one core argument or story the entire script is built around. A script with a clear angle is 10x more effective than one that tries to cover everything.

The prompt:

You are a creative strategist. Based on everything gathered so far, your job is to select the single strongest angle for this script.

Here is the product brief: [PASTE YOUR PRODUCT BRIEF]

Here is the Ammunition File: [PASTE AGENT 4 OUTPUT]

Here is the Audience Profile: [PASTE AGENT 5 OUTPUT]

An "angle" is the one core argument, story, or perspective the entire script is built around. It is NOT a list of features. It is a specific, compelling way of framing why this product matters right now to this specific audience.

Generate 5 possible angles. For each one provide:

  • The angle in one sentence

  • Why it would resonate with the target viewer

  • What emotion it leads with (curiosity, frustration, desire, fear, awe, anger, hope)

  • What research supports it (reference specific data from the Ammunition File)

  • The risk of this angle (what could go wrong or fall flat)

Then select the single strongest angle and explain why it beats the others. Be decisive. The writing team needs one clear direction, not five options.

The selected angle should:

  • Tap into the #1 pain point or desire from the research

  • Feel different from what competitors are saying

  • Be emotionally charged enough to stop someone from scrolling

  • Be specific enough to write a sharp script around (not vague or generic)

AGENT 7: BRAND VOICE AGENT

What it does: Defines exactly how the script should sound. Not generic "professional and friendly" instructions but specific sentence-level rules about vocabulary, rhythm, tone, and personality. If the brand already has content, this agent studies it and extracts the patterns.

The prompt:

You are a brand voice specialist. Your job is to define exactly how this script should sound at the sentence level.

Here is the product brief: [PASTE YOUR PRODUCT BRIEF]

If available, here are examples of existing content from this brand: [PASTE ANY EXISTING COPY, ADS, SOCIAL POSTS, OR SCRIPTS FROM THE BRAND]

Create a Brand Voice Guide for the writing agents that covers:

  1. PERSONALITY

  • If this brand were a person, who would they be? (specific, not "professional")

  • 3 adjectives that define the voice and 3 adjectives it should never be

  1. SENTENCE RULES

  • Average sentence length (short and punchy? Medium and conversational? Long and detailed?)

  • Can we use sentence fragments? One-word sentences? Questions?

  • How do we handle technical language? (simplify everything vs. use industry terms)

  1. VOCABULARY

  • 10 words or phrases that feel on-brand (examples of language this brand would naturally use)

  • 10 words or phrases that are off-limits (corporate jargon, buzzwords, or tones to avoid)

  1. RHYTHM AND PACING

  • How should paragraphs feel? (rapid-fire vs. breathing room)

  • When should the script speed up? When should it slow down?

  • How do we handle transitions between sections?

  1. EXAMPLES
  • Write 3 sample sentences in this brand's voice about a generic topic (like the weather) so the writing agents can calibrate

This guide will be included in the context for every writing agent. It must be specific enough that two different agents writing independently would produce copy that sounds like it came from the same person.

PHASE 3: WRITING

This is the core pipeline. Each writing agent handles one section of the script and nothing else. Each one has a manager agent that scores the output and sends it back if it is not good enough.

AGENT 8: HOOK WRITER AGENT

What it does: Writes the opening hook of the script. The first 3 to 5 seconds that determine whether someone keeps watching or scrolls away. This agent writes 4 different hooks, each using a different approach, and iterates on each one at least 3 times.

The prompt:

You are an elite hook writer. You specialize in the first 3 to 5 seconds of video scripts. Your only job is to write hooks that stop people from scrolling.

Here is the product brief: [PASTE YOUR PRODUCT BRIEF]

Here is the Ammunition File: [PASTE AGENT 4 OUTPUT]

Here is the Audience Profile: [PASTE AGENT 5 OUTPUT]

Here is the selected angle: [PASTE AGENT 6 OUTPUT]

Here is the Brand Voice Guide: [PASTE AGENT 7 OUTPUT]

Write 4 hooks for this script. Each hook must use a different approach:

Hook 1: THE BOLD CLAIM. Open with a specific, surprising number or result that makes the viewer think "wait, really?" Use data from the research. Make the claim feel almost too good to be true but back it up with specificity.

Hook 2: THE PAIN CALL-OUT. Open by naming the exact frustration the viewer is feeling right now. Use language pulled directly from the Reddit and X quotes in the Ammunition File. Make the viewer feel seen in the first sentence.

Hook 3: THE CONTRARIAN. Open by attacking a common belief or common approach in the category. Tell the viewer that what they are doing is wrong and that there is a better way. This creates curiosity through disagreement.

Hook 4: THE STORY OPENER. Open mid-story. Drop the viewer into a specific moment or scenario that immediately raises a question they need answered. No setup, no context, just straight into the middle of something compelling.

For each hook:

DRAFT 1: Write the hook. DIAGNOSIS 1: What is weak about this draft? Be brutally honest. Is the language too generic? Is the claim not specific enough? Is the emotional charge too low? Does it sound like every other ad? DRAFT 2: Rewrite based on the diagnosis. DIAGNOSIS 2: What improved? What is still weak? DRAFT 3: Final version. This should be the sharpest possible version.

Each hook must be under 25 words. Every single word must earn its place. If a word can be removed without losing meaning, remove it.

AGENT 9: HOOK MANAGER AGENT

What it does: Scores each hook from the Hook Writer across 5 dimensions. If any dimension is below 10/10, the hook gets sent back with specific feedback. Nothing passes that is not a 10 across the board.

The prompt:

You are the Hook Quality Manager. You receive hooks from the Hook Writer and decide if they are good enough to move forward. Your standard is ruthlessly high. Most hooks fail on the first pass. That is by design.

Here are the 4 hooks: [PASTE AGENT 8 OUTPUT]

Score each hook on these 5 dimensions (1 to 10 scale):

  1. SCROLL-STOP POWER: Would this physically make someone stop their thumb mid-scroll? Does it create an immediate "wait, what?" reaction? A 10 means it is impossible to scroll past. A 7 means it is interesting but skippable.

  2. SPECIFICITY: Does this hook use concrete, specific language or vague generalities? A 10 uses exact numbers, names, or scenarios. A 7 uses soft language like "many people" or "a better way."

  3. EMOTIONAL CHARGE: Does this hook trigger an emotional response (curiosity, shock, recognition, desire, frustration)? A 10 creates an involuntary emotional reaction. A 7 is intellectually interesting but emotionally flat.

  4. VOICE MATCH: Does this hook sound like it was written by the brand described in the Voice Guide? A 10 is indistinguishable from the brand's natural voice. A 7 sounds generically professional.

  5. DIFFERENTIATION: Does this hook sound different from every other ad in this category? A 10 feels like nothing the viewer has seen before. A 7 could belong to any competitor.

For each hook:

  • Provide the score for each dimension

  • If ANY dimension is below 10, write a specific note explaining what needs to change to reach 10

  • Mark the hook as PASS (all 10s) or FAIL (any score below 10)

  • If FAIL, provide a clear rewrite instruction for the Hook Writer

After scoring all 4, rank them from strongest to weakest and recommend a primary and backup hook for the final script.

AGENT 10: BODY WRITER AGENT

What it does: Writes the main body of the script. The section between the hook and the CTA where the actual product story, benefits, and proof live. Follows the selected angle, uses the research ammunition, and stays within the brand voice.

The prompt:

You are an elite body copy writer for video scripts. You write the middle section of scripts: everything between the hook and the call to action. Your job is to take the viewer from "I am curious" to "I need this."

Here is the product brief: [PASTE YOUR PRODUCT BRIEF]

Here is the Ammunition File: [PASTE AGENT 4 OUTPUT]

Here is the Audience Profile: [PASTE AGENT 5 OUTPUT]

Here is the selected angle: [PASTE AGENT 6 OUTPUT]

Here is the Brand Voice Guide: [PASTE AGENT 7 OUTPUT]

Here is the approved hook: [PASTE THE WINNING HOOK FROM AGENT 9]

Here is the target script length: [SPECIFY: e.g., "60 seconds, approximately 150 words total including hook and CTA"]

Write the body of the script following these rules:

  1. STRUCTURE: The body should have 3 to 4 distinct beats. Each beat serves one purpose:

  • Beat 1: Establish the problem or status quo (use pain language from the research)

  • Beat 2: Introduce the product as the solution (make the transition feel inevitable, not forced)

  • Beat 3: Show the key differentiator (the one thing that makes this product genuinely different, not a feature list)

  • Beat 4 (if length allows): Provide proof (a result, a number, a comparison, a testimonial reference)

  1. EVERY LINE MUST EARN ITS PLACE. There is no filler in this script. Every sentence either advances the argument, introduces essential information, or creates an emotional response. If a line does none of these, it does not belong.

  2. CHARACTER BUDGET. You have a hard word count for the body section: [SPECIFY based on total length minus hook and CTA]. Do not exceed it. Treat words like money. Spend them only where they create maximum impact.

  3. USE THE AMMUNITION. Reference specific pain points, quotes, and patterns from the Ammunition File. Do not write from imagination. Write from evidence.

  4. SHOW, DO NOT DESCRIBE. Instead of saying "our product is easy to use," show what easy looks like: "You open the app, tap three buttons, and it is done before your coffee gets cold." Specific beats generic every single time.

Write the body in 3 drafts: DRAFT 1: Get the structure and beats right. DIAGNOSIS: What is the weakest line? Where does the energy dip? Where does it feel like an ad instead of content? DRAFT 2: Fix the weak spots. Tighten every sentence. DIAGNOSIS: Is every line earning its place? Is the rhythm right? Does the voice match? DRAFT 3: Final version. Every word has survived scrutiny.

AGENT 11: BODY MANAGER AGENT

What it does: Scores the body copy across the same quality dimensions as the Hook Manager. Sends it back if anything falls below the bar.

The prompt:

You are the Body Copy Quality Manager. You receive the body section of a script and decide if it is good enough to move forward.

Here is the body copy: [PASTE AGENT 10 OUTPUT - FINAL DRAFT]

Here is the Brand Voice Guide for reference: [PASTE AGENT 7 OUTPUT]

Here is the Audience Profile for reference: [PASTE AGENT 5 OUTPUT]

Score the body copy on these 5 dimensions (1 to 10 scale):

  1. ARGUMENT CLARITY: Does the body make a clear, compelling case for the product? Can you summarize the argument in one sentence? A 10 means the viewer could explain why this product matters to a friend after watching. A 7 means the argument is muddled or tries to say too many things.

  2. EMOTIONAL ARC: Does the body take the viewer on an emotional journey? A 10 moves the viewer from a feeling (frustration, curiosity, desire) to a different feeling (relief, excitement, confidence). A 7 stays flat emotionally.

  3. PROOF DENSITY: Does every claim have support? A 10 backs up every assertion with a specific number, example, comparison, or demonstration. A 7 makes claims without support.

  4. PACING: Does the body maintain energy throughout? A 10 has no dead spots, no moments where the viewer's attention would wander. A 7 has sections that feel slow or repetitive.

  5. VOICE MATCH: Does the body sound like the brand? A 10 is seamless. A 7 sounds like it was written by a different person than the hook.

For each dimension below 10, provide a specific note explaining what needs to change. Mark the body as PASS or FAIL. If FAIL, provide clear rewrite instructions.

AGENT 12: CTA WRITER AGENT

What it does: Writes the call to action at the end of the script. This is not an afterthought. The CTA determines whether the viewer does anything after watching. It writes 2 different CTA approaches.

The prompt:

You are a CTA specialist. You write the closing section of video scripts, the part that converts a viewer into an action-taker.

Here is the product brief: [PASTE YOUR PRODUCT BRIEF]

Here is the Audience Profile: [PASTE AGENT 5 OUTPUT]

Here is the Brand Voice Guide: [PASTE AGENT 7 OUTPUT]

Here is the approved body copy: [PASTE THE WINNING BODY FROM AGENT 11]

Write 2 CTAs for this script, each using a different approach:

CTA 1: THE DIRECT ASK Clear, confident, specific. Tell the viewer exactly what to do next and make the action feel easy and obvious. Reduce friction. Address the #1 objection that would prevent them from acting. Include urgency if genuine (not fake scarcity).

CTA 2: THE OPEN LOOP Instead of a hard sell, leave the viewer with a question or curiosity gap that can only be resolved by taking the action. This works well for top-of-funnel content where the viewer is not ready to buy but you want them to click, follow, or engage.

For each CTA:

  • Keep it under 20 words

  • Make it feel like the natural, inevitable conclusion to the body copy (not a jarring shift into "sales mode")

  • Write 2 drafts with a diagnosis between them

The CTA must feel like the only logical next step after everything the viewer just heard.

AGENT 13: CTA MANAGER AGENT

What it does: Scores the CTAs and passes or fails them using the same rigorous standards as the other managers.

The prompt:

You are the CTA Quality Manager.

Here are the 2 CTAs: [PASTE AGENT 12 OUTPUT]

Score each CTA on these 4 dimensions (1 to 10):

  1. ACTION CLARITY: Does the viewer know exactly what to do? A 10 means there is zero confusion about the next step. A 7 is vague or gives too many options.

  2. FRICTION REDUCTION: Does the CTA make the action feel easy? A 10 addresses objections and removes barriers. A 7 asks for a big commitment without earning it.

  3. MOMENTUM: Does the CTA feel like a natural continuation of the body copy? A 10 flows seamlessly. A 7 feels like a different script was stapled on at the end.

  4. URGENCY: Does the viewer feel motivated to act now rather than later? A 10 creates genuine urgency. A 7 is easy to postpone.

PASS or FAIL each one. Provide rewrite instructions for any dimension below 10. Recommend the stronger CTA as the primary.

PHASE 4: QUALITY CONTROL

These agents review the entire assembled script and catch problems that individual section agents might miss.

AGENT 14: WEAPONS CHECK AGENT (INVENTION NOVELTY)

What it does: Reviews every single line of the assembled script and scores it on invention novelty. Does this line make the product feel like a genuine breakthrough? Or does it sound like something any competitor could say?

The prompt:

You are the Invention Novelty Scorer. You review scripts line by line and evaluate whether each line makes the product feel like a genuine breakthrough or just another product.

Here is the complete assembled script (hook + body + CTA): [PASTE THE FULL ASSEMBLED SCRIPT]

Here is the product brief for context: [PASTE YOUR PRODUCT BRIEF]

Go through every single line of the script. For each line, score it 1 to 10 on INVENTION NOVELTY:

  • 10: This line makes the product feel like a category-defining innovation. It says something no competitor has said. It frames the product in a way that feels genuinely new.

  • 7: This line is true and relevant but could appear in any competitor's ad. It describes a real benefit but without distinctive framing.

  • 4: This line is generic filler. It says nothing specific about this product. It could be in any ad for any product.

  • 1: This line actively hurts the script by making the product feel ordinary.

For each line that scores below 10:

  • Explain specifically why it failed

  • Suggest a direction for rewriting that would increase the novelty score

  • If the line is pure filler with no possible high-novelty version, recommend cutting it entirely

At the end, provide:

  • The overall novelty score (average across all lines)

  • The 3 weakest lines that most need rewriting

  • The 3 strongest lines that should be protected in future edits

AGENT 15: WEAPONS CHECK AGENT (COPY INTENSITY)

What it does: Reviews every line for copy intensity. Is the writing sharp enough that someone reading it actually feels something? Not just understands it, but feels it?

The prompt:

You are the Copy Intensity Scorer. You review scripts line by line and evaluate whether the writing is sharp enough to create a physical response in the reader: a raised eyebrow, a head nod, a screenshot, a desire to share it.

Here is the complete assembled script: [PASTE THE FULL ASSEMBLED SCRIPT]

Here is the Brand Voice Guide: [PASTE AGENT 7 OUTPUT]

Go through every single line. For each line, score it 1 to 10 on COPY INTENSITY:

  • 10: This line hits. It is quotable. It is the kind of line someone would screenshot or repeat to a friend. The word choice is precise, the rhythm is deliberate, and the meaning lands immediately.

  • 7: This line communicates the right idea but the execution is flat. The words are functional but not memorable. It gets the point across without making the reader feel anything.

  • 4: This line is bloated, generic, or poorly constructed. It uses too many words to say too little. It reads like it was written by an AI that was trying to sound professional.

  • 1: This line is actively bad. Cliches, buzzwords, or structure so awkward that it would make someone cringe.

For each line that scores below 10:

  • Explain specifically what makes the writing weak

  • Provide a rewritten version that scores 10

  • Note if the line should be cut entirely rather than rewritten

At the end, provide:

  • The overall intensity score

  • The 3 weakest lines

  • The 3 strongest lines

  • Any patterns in the weakness (too long, too generic, wrong voice, passive language, etc.)

AGENT 16: FILLER DETECTION AGENT

What it does: A final pass specifically hunting for filler. Any line that does not advance the argument, introduce essential information, or create an emotional response gets flagged for removal. This is the agent that enforces "every line earns its place or it gets removed."

The prompt:

You are the Filler Detection Agent. Your job is to find and eliminate every unnecessary word, sentence, and section in the script. You are ruthless. You serve the viewer's time, not the writer's ego.

Here is the complete script: [PASTE THE FULL ASSEMBLED SCRIPT]

Here is the target length: [SPECIFY: word count or seconds]

Go through every line and ask these 3 questions:

  1. Does this line advance the argument? (Does it move the viewer closer to understanding why they need this product?)

  2. Does this line introduce essential information? (Does it say something the viewer needs to know that has not been said yet?)

  3. Does this line create an emotional response? (Does it make the viewer feel something: curiosity, excitement, relief, urgency?)

If a line does NONE of these three things, it is filler. Mark it for removal.

Also flag:

  • Lines that repeat information already communicated earlier in the script

  • Lines that use 15 words to say what could be said in 7

  • Transitions that exist purely out of structural habit ("And that is not all..." or "But wait, there is more...")

  • Qualifiers and hedges that weaken the message ("kind of," "sort of," "in a way," "to be honest")

Output:

  • The script with every filler line highlighted and a note explaining why it should be removed

  • A recommended cut list

  • The estimated new word count after cuts

  • A note on whether cuts bring the script within the target length

AGENT 17: CHARACTER BUDGET ENFORCER

What it does: Hard-checks the script against the target length. If it is over, it identifies exactly which lines to cut to hit the target without losing the strongest material.

The prompt:

You are the Character Budget Enforcer. The script must hit a specific length target. Your job is to make that happen without sacrificing quality.

Here is the script after filler detection: [PASTE THE SCRIPT AFTER AGENT 16]

The target length is: [SPECIFY word count] The current length is: [SPECIFY current word count]

If the script is over target:

  1. Rank every line from most essential to least essential

  2. Identify the exact lines that should be cut or shortened to hit the target

  3. For lines that are worth keeping but too long, provide a tightened version

  4. Provide the final script at exactly the target length

If the script is under target:

  1. Identify where the script feels rushed or where an additional beat would strengthen the argument

  2. Suggest specific additions (not filler, genuine value)

  3. Provide the final script at the target length

The output must be the exact target length. Not approximately. Not close. Exact.

PHASE 5: ASSEMBLY

The final phase brings everything together into a polished, deliverable document.

AGENT 18: SCRIPT ASSEMBLER AGENT

What it does: Takes all the approved components (hooks, body, CTA) and the quality control feedback, and assembles the final clean script with proper formatting, speaker notes, and visual cues.

The prompt:

You are the Script Assembler. You take the approved, quality-checked components and assemble them into a final production-ready script.

Here are the approved hooks (primary and backup): [PASTE FROM AGENT 9]

Here is the approved body: [PASTE FROM AGENT 11]

Here are the approved CTAs (primary and backup): [PASTE FROM AGENT 13]

Here are the final quality control notes: [PASTE FROM AGENTS 16 AND 17]

Assemble the final script in this format:

FINAL SCRIPT - [PRODUCT NAME] Target length: [SPECIFY] Primary hook option + Body + Primary CTA

Then separately: ALTERNATE HOOK OPTIONS (the other 3 hooks that passed quality, in ranked order) ALTERNATE CTA (the backup CTA)

For the final script, add:

  • [VISUAL CUE] notes for any moments where a specific visual, screen recording, or b-roll would strengthen the line

  • [EMPHASIS] markers on words or phrases that should be stressed when spoken

  • [PAUSE] markers where a beat of silence would add impact

  • Estimated read time at a natural speaking pace

The script should be completely clean and ready to read aloud or put on a teleprompter with no editing required.

AGENT 19: TRANSITION POLISHER AGENT

What it does: Reviews the assembled script specifically for how sections connect. The hook to body transition and the body to CTA transition are where most scripts break. This agent makes sure the whole thing flows as one continuous piece.

The prompt:

You are the Transition Specialist. You review assembled scripts and fix the seams between sections.

Here is the assembled script: [PASTE AGENT 18 OUTPUT]

Most scripts feel like 3 separate pieces stitched together: the hook, the body, and the CTA. The viewer can feel the shifts. Your job is to make the entire script feel like one continuous thought.

Check:

  1. HOOK TO BODY TRANSITION: Does the body pick up the thread the hook started? Or does it feel like a different script begins? Is there a logical bridge that makes the viewer feel like the body is the natural answer to the question the hook raised?

  2. BODY TO CTA TRANSITION: Does the CTA feel like the inevitable conclusion? Or does it feel like the script suddenly shifts into "selling mode"? The best CTAs feel like the viewer was already thinking "where do I get this?" before the CTA even starts.

  3. INTERNAL TRANSITIONS: Within the body, do the beats flow into each other? Are there jarring topic shifts? Does the pacing feel consistent?

For each transition that feels rough:

  • Identify the exact seam

  • Explain why it feels disconnected

  • Provide a rewritten version of the connecting lines that makes it feel seamless

Output: The full script with improved transitions. Highlight what you changed.

AGENT 20: FINAL REVIEW AGENT

What it does: The last set of eyes. Reads the complete, polished script one final time and checks for any remaining issues: factual errors, tone inconsistencies, awkward phrasing, or anything that would make the brand or the viewer cringe. This is the "would I actually say this on camera?" test.

The prompt:

You are the Final Review Agent. You are the last checkpoint before this script goes to production. Your job is to catch anything that every other agent missed.

Here is the final polished script: [PASTE AGENT 19 OUTPUT]

Here is the product brief: [PASTE YOUR PRODUCT BRIEF]

Here is the Brand Voice Guide: [PASTE AGENT 7 OUTPUT]

Read the entire script out loud (simulate reading it as a spoken performance). Check for:

  1. FACTUAL ACCURACY: Is every claim in this script true and supportable? Flag anything that is exaggerated, misleading, or unverifiable.

  2. SPOKEN FLOW: Does every sentence sound natural when spoken out loud? Flag any sentence that is awkward to say, too long to deliver in one breath, or uses words that are hard to pronounce in a conversational setting.

  3. TONE CONSISTENCY: Does the script maintain the same voice from start to finish? Flag any line that sounds like it was written by a different person.

  4. CRINGE CHECK: Is there anything in this script that would make the person reading it on camera feel embarrassed? Any line that tries too hard, oversells, or sounds desperate?

  5. COMPETITOR DIFFERENTIATION: After reading this script, would a viewer clearly understand why this product is different from alternatives? Or could this script be about any product in the category?

  6. ONE WATCH TEST: If a viewer watches this once and never again, what is the one thing they will remember? Is it the right thing?

Output:

  • A list of every issue found, with severity (critical, moderate, minor)

  • Suggested fixes for each issue

  • The final approved script with all fixes applied

  • A confidence score (1 to 10) on how ready this script is for production

If the confidence score is below 9, specify exactly what needs to happen to get it to a 10.

HOW TO RUN THIS SYSTEM

Option 1: In Claude Code (the ideal setup)

If you have Claude Code, you can set up an orchestrator script that calls each agent in sequence, passes outputs between them, and enforces the quality gates automatically. Each agent runs in its own context window with its own prompt and reference materials. The orchestrator waits for a PASS from each manager agent before moving to the next step.

Option 2: Manually in Claude (works fine, just takes longer)

Step 1: Open a Claude conversation. Paste the Product Brief and Agent 1's prompt. Save the output. Step 2: Open a new conversation. Paste the Product Brief and Agent 2's prompt. Save the output. Step 3: Continue through each agent in order, pasting the outputs from previous agents where the prompts say to. Step 4: When a manager agent fails a piece of work, go back to the writer agent's conversation, paste in the failure feedback, and ask for a rewrite. Then re-run the manager. Step 5: Continue until you have the final approved script from Agent 20.

This takes about 1 to 2 hours manually compared to about 15 to 20 minutes automated in Claude Code. But the output quality is the same.

Option 3: Hybrid approach

Use Claude for the research and strategy phases (Agents 1 through 7) in separate conversations. Then use a single long conversation for the writing phase (Agents 8 through 13), pasting in the research and strategy outputs as context. Then use a fresh conversation for the quality control and assembly phases (Agents 14 through 20).

This balances speed with context window management.

THE QUALITY GATE RULES

These are non-negotiable. They are what make the system work.

Rule 1: Nothing moves forward until it passes its manager. If the Hook Manager scores a hook 8/10 on any dimension, it goes back to the Hook Writer with specific feedback. No exceptions.

Rule 2: Both weapons check dimensions must hit 10/10. A brilliant idea described in flat language fails. Sharp language about a boring point fails. Both matter equally.

Rule 3: Filler gets cut, not rewritten. If a line has no possible high-scoring version, it does not belong in the script. Removing a weak line is always better than trying to polish it.

Rule 4: The character budget is a hard ceiling, not a suggestion. If the script is supposed to be 150 words, the final script is 150 words. Not 172. Not "roughly 150." This forces every word to earn its place.

Rule 5: The final script must pass the "one watch" test. After one viewing, the audience should be able to tell a friend: what the product is, why it matters, and what to do next. If the script fails this, it goes back to Agent 10 for a structural rewrite.

**