CodeWords: the UGC engine case study

Real product data

Where Hook fits: the research stage

Before templatizing a format, Hook scanned competitor and creator accounts in the target niche and scored each video on real virality, not raw view count.

The bar: don’t templatize off one outlier. Wait for more than one video independently outperforming in the same shape. Everything past this point, script generation, sourcing, scheduling, is CodeWords’ own system, included because it’s what turns a research signal into shipped video, not because it runs on Hook.

What Hook actually is: a workspace that watches the organic and paid video in a category, separates real virality from boosted spend, and surfaces the formats worth building around, the exact job it did in stage one here.

1 · Overview

The full system, six stages

Can an AI agent run most of a paid-UGC operation, with a human only making calls that need taste? CodeWords lands at roughly 70-80% hands-off, by design, section 10 covers the rest honestly.

The stack, one of six tools is Hook

The six stages

Why it compounds

A winning script becomes the next research reference, so each cycle starts from a stronger baseline, and that rising bar is exactly what Hook is checking for.

Building your own version

1. Connect Hook, Notion, and Buffer via MCP.
2. Write a CLAUDE.md describing your product and ICP. Template in section 4.
3. Budget enough to seed a few sub-accounts through their first batch. Section 9.
4. Pick a format from Hook’s Discover feed and run stages 1-3 once before sourcing a creator.

2 · Setup

Tools, accounts, and setup

MCP wiring order

1. Hook first. Add competitor and creator accounts, a manual pass the first time.
2. Notion second. One page, sub-pages per format, is enough to start.
3. Buffer last, it depends on social accounts already existing. See section 6.
4. Claude drives Fiverr and Drive directly through your logged-in browser, no separate MCP needed.

Connecting Hook’s MCP

Nothing to install or mint. In Claude, Settings → Connectors → Add custom connector, and paste:

https://api.askhook.com/api/mcp

Same URL works in Cursor’s .cursor/mcp.json or any other MCP client, sign in to Hook when prompted on first connect.

Full walkthrough: askhook.com/mcp

CLAUDE.md: company context, once

A minimal starting version:

# CLAUDE.md

## Company
[Your company] builds [what it does] for [who it's for].

## Product
[One paragraph: what the product actually does, in plain language.
This is what Claude will describe in every script.]

## ICP / industries we sell into
[List the industries, use cases, or personas you want script variants for.
This becomes the variation axis in stage 2.]

## Claims we can make
[What you can honestly say the product does, and anything you must NOT
claim. This is the single most important section, see section 4.]

Cost floor

Budget at least $2,000 for a real first pass, enough for 3 sub-accounts through their first batches. Ad spend is separate and optional. Full ladder in section 9.

Minimum viable version

• –Hook's lowest tier: browse the feed manually, skip MCP automation
• –Skip Fiverr for a first test, film a variant in-house
• –Skip Buffer's paid tier while running one account
• –Skip ElevenLabs, paid analytics, and all ad spend

3 · Brief template

The Notion brief template (“Format Pack”)

One page per format, same skeleton every time.

4 · Prompts

The actual prompts

Most of the work is context Claude already has, the prompt itself is short.

Prompt 1: templatize a winning format

Templatize this format. Break it into its fixed beats (the parts that
never change) and its variable slots (the parts that change per
industry/use case/persona). Then generate <N> script variants, one per
<industry|use case|persona> from the list in CLAUDE.md, keeping every
fixed beat identical and only changing the variable slots.

Create a Notion page for this using the Format Pack template, and put
the scripts in packs of 10.

This works with almost no extra detail because Claude already has the reference video, the CLAUDE.md, and the template. Thin any one of those, and the output turns generic.

Worked example

Reference: a DM-reveal hook (low number typed, client reveals a bigger budget, creator counters higher), then a build sequence, then a payoff.

Templatize the Budget Reveal format from this video. The DM hook and
the numbers in it stay identical across every variant, don't touch
that part. Vary only the build sequence: swap the industry the client
is in and what gets built for them, using the industries in CLAUDE.md.
Generate 10 variants, one pack, Aesthetics through Gym.

Output: 10 scripts, identical DM dialogue, a different industry’s build sequence each time.

The variation axes

• –Industry, CodeWords’ main axis
• –Use case, multiple jobs in one industry
• –Persona, different buyer types
• –Feature, a different capability per variant

The failure mode to watch for

A lazy variant: same pain point, only the industry noun swapped. Spot-check a few from different ends of the batch. Second check, every batch: does every claim still hold up against CLAUDE.md’s claims list?

5 · Creators

Finding, vetting, and QA-ing creators

A creator, mid-hook

Claude drives a logged-in Fiverr session: searches the UGC-video category, filters to price, screens for camera presence, not spoken fluency. Baseline: £20-30 per video.

Language usually isn’t a filter, non-native speakers who act well are the cheapest option. If a format needs spoken voice, shoot silent b-roll and layer an ElevenLabs voiceover in instead of dropping the creator.

What disqualifies a creator

• –Inconsistent presentation across their own samples, a reseller-account tell
• –Poor audio or lighting, even at the right price
• –No natural camera reaction, some formats need it even without dialogue

QA gates

Two are hard gates, fail either and the video is rejected regardless of everything else.

6 · Accounts

Account warm-up

A new account posting immediately, especially alongside others sharing signals (device, network), reads like a bot farm. Patience before the first post is the fix.

The sequence

Mechanics

• –A distinct phone number per account (eSIM or similar)
• –A distinct email alias, ideally on your own domain
• –A real-feeling profile: genuine photo, matching bio

The real failure mode is a temporary checkpoint, not a permanent ban, budget for it as routine.

7 · Audio

Adding soundtracks programmatically

Silent b-roll and assembled clips need music muxed in without a human touching an editor.

1. Detect if a clip needs music

ffmpeg -i clip.mp4 -af volumedetect -f null - 2>&1 | grep mean_volume

-90dB or lower is effectively silent.

2. The rule

A track must be at least as long as the clip. It plays once, trimmed, never looped, a loop point reads as cheap immediately.

3. The mux command

ffmpeg -i clip.mp4 -ss <track_offset> -i track.mp3 \
  -map 0:v:0 -map 1:a:0 \
  -c:v copy -c:a aac \
  -shortest \
  out.mp4

-shortest enforces the trim.

The copyright tradeoff

Commercial tracks risk a platform muting the post’s audio. Build the library from royalty-free or explicitly licensed sources.

8 · Scheduling

Scheduling and publishing: Buffer

Buffer schedules and reports real per-post metrics. Every rule below cost a broken queue before it became a rule.

A finished, scheduled video

createPost(input:{
  channelId, schedulingType:"automatic", mode:"customScheduled", dueAt:"<ISO-8601 UTC>",
  text, metadata:{ instagram:{ type:"reel", shouldShareToFeed:true } },
  assets:[{ video:{ url } }]
})

Six rules that break a queue silently

1. schedulingType is fixed at creation. Always "automatic", wrong ones get recreated, not patched.
2. The media URL must stay alive past publish time, not just scheduling time.
3. HEVC/H.265 files (iPhone footage) need transcoding, or ask for H.264 directly.
4. Rate limits trip mid-batch around 50 rapid calls.
5. editPost replaces the whole post, resend everything, even for a schedule-only edit.
6. Delivered clips are often silent, check, don’t assume.

9 · Economics

Economics

The cost ladder

What actually drives cost

• –Boosting, not creator cost, is usually the largest recurring line
• –Account launch cost is fixed and one-time per account
• –Production, not posting, is the real bottleneck

10 · The honest part

What this system will not do for you

A mature run is roughly 70-80% hands-off. Every batch still needs a manual pass on QA (“does this feel authentic” stays human) and occasional platform repair work, automation reduces both, it doesn’t erase them.

Three decisions that stay human

The system removes the repetitive 80%. The 20% that stays manual is where a wrong automated call is expensive, and where judgment doesn’t reduce to a rule.

FAQ

Common questions