Attract Signal

Overview

This skill turns short-form content sources into reusable content-intelligence briefs. It supports YouTube Shorts, TikTok, Instagram Reels, X video, and other short-form exports. It is designed for competitive/trend research where the user wants to find high-performing short-form content, understand why it worked, and translate the underlying signal into original brand strategy without copying the original. It is industry-agnostic by default: do not assume tattoo, beauty, SaaS, local services, restaurants, ecommerce, coaching, fitness, or any other niche unless the user provides that context.

This is a portable skill for Hermes, Codex, and Claude. Keep the workflow agent-neutral: local Markdown and CSV outputs are universal; Google Docs/Sheets publishing is an optional enhancement when gogcli is installed and authenticated.

Default example channel for testing:

https://www.youtube.com/@_The_Clean_Girl/shorts

The standard threshold is 10,000+ likes, but the user can change it. Always cite source Shorts links in the output.

When to Use

Use this skill when the user asks to:

Scan one or more short-form sources for winners, including YouTube Shorts channels and normalized TikTok, Instagram Reels, X video, or platform exports.
Filter Shorts by likes/views/engagement.
Analyze hooks, visual hooks, formats, content trends, and transcript patterns.
Create a signal and trend breakdown that can become original scripts, shot lists, Google Docs, or storyboards.
Build a source-cited content signal library.
Create an industry-agnostic or brand-specific 14-day content testing sprint.

Don't use this for long-form YouTube summaries only; use youtube-content directly for single-video transcript transforms.

Related Skills / Tools

Bundled scripts/fetch_transcript.py — fetches transcripts from individual Shorts or videos.
youtube-content — optional fallback transcript skill if already installed.
gogcli — preferred Google CLI for writing the default Google Docs copy (brew install openclaw/tap/gogcli).
google-workspace — optional Google Workspace fallback if gog/gogcli is not installed or not authenticated and the host agent provides that skill/tool.
image_generate tool — use later to generate storyboard frames after the script/shot list is approved.
last30days — preferred independent evidence collector for current Reddit threads and comments. Keep it separately installed and updateable; Attract Signal consumes its output instead of vendoring its engine.

Deep Reddit Research Mode

Use this mode when the user wants Reddit audience research, voice-of-customer language, content gaps, or current conversations translated into content strategy.

Run the independently installed last30days skill for the topic and include Reddit. Follow its own setup, source-health, research, and citation contract exactly.
Locate the raw Markdown artifact from the Last30Days footer, or use normalized Reddit JSON containing titles, bodies, communities, dates, engagement, comments, and URLs.
Classify and rank the evidence:

python3 $SKILL_DIR/scripts/analyze_reddit_conversations.py \
  ~/Documents/Last30Days/<topic>-raw.md \
  --topic "<topic>" \
  --out reddit-conversation-signals.json \
  --markdown reddit-conversation-signals.md

Read references/reddit-deep-research-prompt.md completely before semantic synthesis. Use it to turn the analyzer output into audience findings, content gaps, hooks, scripts, shot lists, and 14-day sprint inputs.
Preserve each Reddit URL beside every finding and content angle. Never invent a quote, metric, price, date, subreddit-growth claim, or demand signal. Do not treat complaints alone as purchase intent.

The five conversation lenses are multi-label:

Pain Points
Solution Requests
Money Talk
Hot Discussions
Seeking Alternatives

Keep the integration one-way and update-safe: Last30Days owns discovery/freshness; Attract Signal owns conversation analysis and content transformation. Do not copy the Last30Days engine into this repository.

Setup

Install metadata/transcript dependencies if missing:

python3 -m pip install -U yt-dlp youtube-transcript-api

Default Google Docs backend for delivery:

brew install openclaw/tap/gogcli
gog --version

If Homebrew is unavailable, use a host-agent Google Workspace tool or Docker install path from the gogcli skill when available.

The default report workflow writes a local Markdown file first, then creates a Google Doc copy with gog docs create --file. Reports should normalize every channel into Avatar, Promise, Proof, and Path, then build Hooks -> Meats -> CTAs -> 14-day tests. If gog is missing or unauthenticated, keep the local Markdown and tell the user exactly what failed. Use --no-google-doc only for local-only tests or CI.

For shell examples, resolve the installed skill path once:

# Hermes
export SKILL_DIR="${HERMES_HOME:-$HOME/.hermes}/skills/attract-signal"

# Codex
export SKILL_DIR="${CODEX_HOME:-$HOME/.codex}/skills/attract-signal"

# Claude Code
export SKILL_DIR="$HOME/.claude/skills/attract-signal"

Workflow

1. Confirm source and target inputs

Ask for these inputs when they are missing:

Source inspiration links:
- YouTube Shorts channel URL
- Instagram/Reels profile URL
- TikTok profile URL
- X/video profile URL

Target brand links:
- brand website URL
- product page URL
- Instagram/TikTok/YouTube accounts if available

Then ask only the minimum strategy intake needed to set defaults:

What is your main website or store URL? Optional, but if you skip this I will use more generic language.
Which best describes you: creator, product_brand, service, b2b_saas, education, or other?
In one sentence, what are you trying to get viewers to do or get?
Who is this channel mainly for?
If I cannot infer it: do you prefer face_led, product_led, or faceless videos?

Map business_type to internal defaults when the user does not provide explicit overrides:

business_type	default path	default proof/meats	default style
`creator`	`sub`	Story + Demonstration	`face_led`
`product_brand`	`click`	Demonstration + Testimonial	`product_led`
`service`	`book_call`	Demonstration + Testimonial	`face_led`
`b2b_saas`	`book_call`	Demonstration + Education	`face_led`
`education`	`opt_in`	Education + Story	`face_led`

The shortest acceptable prompt is:

Share a source YouTube/Instagram/TikTok/X channel to analyze, plus your brand website or product URL if you want a strategy for your own brand.

Default scan values:

channel_shorts_url = https://www.youtube.com/@_The_Clean_Girl/shorts
min_likes = 10000
max_videos = 50
sort = channel/default order unless user says newest/popular
target_brand_url = optional, but required for product-mode claims unless brand.yaml supplies verified product details

If the user provides only a source channel, generate a generic/creator-style signal report. If the user provides a brand website or product page, open/research it first, then populate brand.yaml fields from verified page text. Never infer product names, offers, discount codes, ingredients, safety claims, or prices without a provided/verified source URL.

2. Collect Shorts metadata

Use the helper script in this skill:

python3 $SKILL_DIR/scripts/scan_shorts.py \
  "https://www.youtube.com/@_The_Clean_Girl/shorts" \
  --min-likes 10000 \
  --max-videos 50 \
  --out ~/youtube-shorts-research/clean-girl-shorts.json \
  --markdown ~/youtube-shorts-research/clean-girl-shorts.md

If YouTube returns Sign in to confirm you’re not a bot, rerun with one of:

python3 $SKILL_DIR/scripts/scan_shorts.py \
  "https://www.youtube.com/@_The_Clean_Girl/shorts" \
  --min-likes 10000 \
  --max-videos 50 \
  --cookies-from-browser chrome \
  --out ~/youtube-shorts-research/clean-girl-shorts.json \
  --markdown ~/youtube-shorts-research/clean-girl-shorts.md

python3 $SKILL_DIR/scripts/scan_shorts.py \
  "https://www.youtube.com/@_The_Clean_Girl/shorts" \
  --min-likes 10000 \
  --max-videos 50 \
  --cookies /path/to/youtube-cookies.txt \
  --out ~/youtube-shorts-research/clean-girl-shorts.json \
  --markdown ~/youtube-shorts-research/clean-girl-shorts.md

The script uses yt-dlp to read channel Shorts and per-video metadata. It also adds channel baseline stats, normalized engagement, relative performance, signal_score, and signal_reason. YouTube may hide likes or throttle metadata; if like_count is missing for many videos, do one of:

Retry with fewer videos (--max-videos 20).
Retry with --cookies-from-browser chrome or a Netscape cookie file if YouTube asks for sign-in/bot confirmation.
Use YouTube Data API directly if the user has a key/OAuth, but first verify the local CLI surface (gog schema / gog --help) before assuming a gog yt command exists. Note: YouTube Data API gives views/comments; public like counts may be unavailable depending on API behavior/privacy.
Fall back to view threshold and mark likes as unavailable.

3. Fetch transcripts for shortlisted Shorts

For each source URL selected by the scanner, use this skill's bundled transcript script:

python3 $SKILL_DIR/scripts/fetch_transcript.py "SHORTS_URL" --timestamps

If this skill is installed without the bundled script for some reason, use the host agent's YouTube transcript skill/tool as a fallback when available.

If transcript is unavailable:

Try again with no language preference.
Use title, description, visible captions, and audio/visual observation if the user provides video access.
Mark transcript_status: unavailable instead of inventing words.

4. Combine multi-channel signals when needed

If the user gives multiple channels, run one scan per channel, then combine them:

python3 $SKILL_DIR/scripts/analyze_signals.py \
  channel-1.json channel-2.json \
  --out signals.json \
  --top 25

The combined output ranks signals across channels with cross_channel_signal_score and preserves every source_url.

4b. Import non-YouTube platforms from exports

YouTube Shorts is the only built-in live scraper. For TikTok, Instagram Reels, X video, or other platforms, normalize user-provided CSV/JSON exports:

python3 $SKILL_DIR/scripts/import_platform.py \
  --platform mixed \
  --input platform-export.csv \
  --source-name "Competitor multi-platform export" \
  --out platform-normalized.json

Then include platform-normalized.json in analyze_signals.py alongside YouTube scans. This gives cross-platform normalization and platform-specific strategy recommendations without pretending to scrape restricted platforms.

5. Generate the strategy report

For generic industry-agnostic strategy:

python3 $SKILL_DIR/scripts/generate_report.py \
  --signals signals.json \
  --transcripts-dir transcripts \
  --out attract-signal-report.md \
  --calendar content-sprint.csv

This creates both:

attract-signal-report.md locally
a native Google Doc copy, with metadata saved as attract-signal-report.google-doc.json
content-sprint.csv, a 14-day test sprint matrix

For brand-specific strategy, pass a simple brand.yaml:

brand_name: Example Brand
business_type: service
brand_url: https://example.com
industry: local service business
audience: busy homeowners who want trustworthy help
offer: a clear, reliable service package
tone: helpful, direct, practical, and warm
proof_points:
  - before-and-after results
primary_path: ""
channel_style: ""
product_mode: ""
product_name: ""
product_url: ""
discount_code: ""
constraints:
  - film with a phone
filming_resources:
  - owner on camera
forbidden_claims:
  - guaranteed results

Then run:

python3 $SKILL_DIR/scripts/generate_report.py \
  --signals signals.json \
  --brand brand.yaml \
  --transcripts-dir transcripts \
  --out attract-signal-report.md \
  --calendar content-sprint.csv \
  --doc-title "Attract Signal Brief - <Brand>"

For local-only testing, add:

--no-google-doc

Transcript files should be named <video_id>.json and generated with:

python3 $SKILL_DIR/scripts/fetch_transcript.py "SHORTS_URL" --timestamps > transcripts/VIDEO_ID.json

6. Analyze each winning Short

For every Short above threshold, produce this breakdown:

## Source: <title>
- Source link: <URL>
- Channel: <channel>
- Published: <date if available>
- Views: <view_count or unknown>
- Likes: <like_count or unknown>
- Comments: <comment_count or unknown>
- Engagement notes: <likes/views ratio if both known; otherwise what is known>

### Transcript / Spoken Structure
- 0-2s: <opening line or caption>
- 2-5s: <setup/escalation>
- 5-12s: <payoff/demo/proof>
- CTA/end: <ending line or loop>

### Trend Type
Choose one or more:
- transformation / before-after
- satisfying process / cleaning reset
- problem-solution
- product demo / tool reveal
- routine / ritual
- aesthetic aspiration
- myth-busting / mistake correction
- listicle / quick tips
- storytime / confession
- challenge / comparison
- social proof / results
- other: <name it>

### Hook
- Verbal hook: <first sentence/caption/promise>
- Curiosity gap: <what question it creates>
- Emotional trigger: <relief, disgust, aspiration, surprise, urgency, etc.>
- Specificity: <numbers, time, room/product/problem named>

### Visual Hooks
- First frame: <what viewer sees immediately>
- Motion: <scrub, pour, wipe, reveal, hand entering frame, jump cut, etc.>
- Contrast: <dirty/clean, clutter/empty, dull/shiny, chaos/order>
- Text overlay: <exact or summarized overlay>
- Pattern interrupt: <unexpected object, speed ramp, close-up, sound sync>

### Signal Pattern
- Core mechanic: <repeatable content formula>
- Why it likely worked: <viewer psychology>
- Original-brand angle: <how to adapt without copying>
- Avoid copying: <what not to reuse verbatim>

7. Synthesize channel-level trends

After individual analyses, make a channel/trend brief:

# Short-Form Content Signal Brief: <source/channel>

## Strategy Spine
- Avatar: <primary viewer and optional secondary viewer>
- Promise: <why the channel exists in one sentence>
- Proof: <two dominant meats: Demonstration, Testimonial, Education, Story>
- Path: <sub, click, opt_in, buy, apply>
- Channel style: <face_led, product_led, faceless>

## Top Winners
| Rank | Source | Views | Likes | Trend type | Hook type | Strategy angle |
|---|---:|---:|---|---|---|---|

## Hook Library
- Top 5 raw hooks:
- Generalized hook templates:
- 2 ready-to-read hook lines for each top template:
- Winner-adjacent variations:
- Pattern tags: Curiosity, Challenge, Spectacle, Transformation, Social Proof, Narrative
- If the top 5 collapse into the same pattern/template, branch the dominant pattern into 3-5 sub-templates using title/transcript cues, then rotate those sub-templates through the 14-day sprint.
- For comedy/open-mic reports, turn Narrative collapse into joke sub-templates such as Shock Answer, Roast Escalation, Underdog Reversal, Instant Character, Backfire Bit, Crowd-Work Premise, and Tag Ladder.

## Repeating Visual Patterns
- Visual formulas:
- Editing rhythm:
- Common props/products/settings:
- Common emotions:
- Common CTAs/loop endings:

## Meats, Style, And CTA System
- Primary meats:
- Channel style adjustment:
- Product mode, when click/buy or product-led:
  - show product
  - show application/use
  - show result
- Product names, offer claims, URLs, and discount codes must come from `brand.yaml` or a verified product page. Do not invent product names, fake brands, fake coupon codes, or unsupported claims.
- For comedy/open-mic reports, convert `Story` meat into a bit skeleton: setup, assumption, turn, tag.
- Short CTA variants by path, with IDs like `cta_sub_1` or `cta_apply_1`:

## Script Drafts
For each concept:
- Title / working caption
- Hook from library
- Meat
- Payoff
- CTA
- On-screen text
- Voiceover/dialogue
- Source inspiration links

## 14-Day Sprint Matrix
| Day | Test mix | Hook template | Script line 0-2s | Pattern tag | Meat | Product step | Style ID | CTA variant ID | Source |
|---:|---|---|---|---|---|---|---|---|---|

## Shot Lists
For each concept:
| Shot | Duration | Framing | Action | Props | Overlay | Notes |

## Storyboard Prompts
For each key frame:
- Frame prompt for image generation
- Camera/framing
- Lighting/style
- Brand notes
- Source reference link(s)

8. Default delivery: local Markdown plus Google Doc

The expected final deliverable is both a local Markdown file and a Google Doc copy containing the trend brief, citations, scripts, shot lists, and storyboard prompts. Draft locally first as Markdown; the report generator now publishes that Markdown to Google Docs by default when gogcli is installed and authenticated.

Preferred report-and-publish helper:

python3 $SKILL_DIR/scripts/generate_report.py \
  --signals signals.json \
  --brand brand.yaml \
  --transcripts-dir transcripts \
  --out attract-signal-report.md \
  --calendar content-sprint.csv \
  --doc-title "Attract Signal Brief - <Brand or Channel>"

Use --open-doc when the user explicitly wants the generated Doc opened in the browser. The standalone publisher remains available for already-generated Markdown:

python3 $SKILL_DIR/scripts/publish_doc.py \
  --file attract-signal-report.md \
  --title "Attract Signal Brief - <Brand>"

Direct gogcli flow for the current gog CLI:

gog docs create "Short-Form Signal Brief - <Channel>" --file brief.md --json
# parse the returned doc ID, then verify the created file:
gog drive get <docId> --json --select id,name,mimeType,webViewLink,owners

If a blank Doc already exists and the agent needs to append a local Markdown draft:

gog docs write <docId> --append --file brief.md --json
# verify the created file:
gog drive get <docId> --json --select id,name,mimeType,webViewLink,owners

If gog is unavailable, keep the local Markdown artifact and tell the user exactly what OAuth/install step is missing; use a host-agent Google Workspace flow only when explicitly requested and available. Never share, permission-change, or overwrite Google Docs without user approval.

9. Google Sheets sprint export

After the user approves a Sheets write:

python3 $SKILL_DIR/scripts/export_calendar_sheets.py \
  --calendar content-sprint.csv \
  --title "Attract Signal Sprint - <Brand>"

10. Reusable signal library

Save top signals for compounding strategy memory:

python3 $SKILL_DIR/scripts/signal_library.py add --signals signals.json --top-only
python3 $SKILL_DIR/scripts/signal_library.py list
python3 $SKILL_DIR/scripts/signal_library.py search "challenge"

Ethical / Brand Safety Rules

Extract patterns; do not copy exact scripts, claims, edit sequences, or distinctive creative expressions.
Keep citations beside every source-inspired idea.
Rewrite into the user's brand voice, products, audience, and proof points.
If a source makes a claim, do not repeat it as factual for the user's brand unless the user can substantiate it.
Do not download or reuse creator footage/assets unless the user has rights.

Common Pitfalls

Likes are hidden or missing. Mark them unknown and filter by available metrics instead of fabricating likes.
Shorts tab pagination can be flaky. Retry with lower --max-videos, update yt-dlp, or use browser/manual sampling for the first page.
Metadata-only extraction can hit format errors. The scanner uses --ignore-no-formats-error for per-video JSON so unavailable video formats do not block metadata collection.
Transcripts often fail for Shorts. Use available captions only; otherwise describe visual/audio structure from metadata or browser observation and label transcript as unavailable.
Confusing inspiration with copying. Always translate source patterns into new concepts and cite references.
Skipping the local artifact. Always write the local Markdown first; the Google Doc is a copy of that source file.

1GMedia/attract-signal-skill

attract-signal-skill 是什么？

在你喜欢的 AI 中提问

文档