Customer Research

Customer Research

Guide for gathering and synthesizing real customer intelligence — from online communities, review sites, video comments, and social platforms — using the Hyper MCP scraper toolkit.

The goal is always the same: surface what customers actually say (in their own words), not what you assume they say.

Out of scope — defer to other skills

Requirements

• Hyper MCP installed. https://app.hyperfx.ai/mcp
• Reddit Scraper, Twitter Scraper, YouTube Scraper, TikTok Scraper and Instagram Scraper toolkits enabled at https://app.hyperfx.ai/apps.

Not all scrapers need to be active for every run — enable the ones relevant to your ICP (Reddit and one review site is the minimum). If a scraper tool is not found, skip that source and continue with the others.

How to run the tools in this skill

Every tool in this skill is named by its canonical tool name. Run it with the call your surface gives you:

If a tool is not found: in Hyper chat, find its toolkit id with discover_toolkits and pass it to enable_toolkit. On an MCP client or the CLI, tell the user which integration to connect in Hyper.

Tool surface

Critical rules

1. Always capture verbatim language. Don't paraphrase customer quotes — the exact words are what gets used in copy and messaging. Extract and preserve them.
2. Scrape before summarizing. Don't rely on your training data to describe what customers say about a product. Actually fetch the sources.
3. Label confidence on every insight. High = 3+ independent sources, unprompted. Medium = 2 sources or prompted only. Low = single source. Never present a Low-confidence finding as a conclusion.
4. Mind the bias of each source. Reddit skews technical and skeptical. Review sites skew toward power users and people with strong opinions. Support tickets skew toward problems. Factor this in before generalizing.
5. Don't invent persona details. If you don't have data for a persona field, leave it blank rather than filling it in with assumptions.
6. youtube_video_transcripts_fetch is slow (~15–30s). It spins up an isolated sandbox. Only use it for videos where the language in the spoken content (not comments) is what matters.

Two modes

Most research combines both modes. Establish which applies before starting.

Mode 1 — Analyze existing assets

The user provides raw material: interview transcripts, survey responses, NPS verbatims, support tickets, win/loss notes. No tool calls needed — the job is extraction and synthesis.

Read references/synthesis-templates.md for the extraction framework, persona template, and VOC quote bank format. Then produce the requested deliverable.

Mode 2 — Go find research online

The user needs intel from online communities, review sites, and social platforms. This is where MCP tools do the heavy lifting.

The steps below name the tool of each source. Read references/source-playbooks.md only when a scrape fails or returns nothing.

Mode 2 workflow

Bias toward action. If the user's message includes a product name (or URL) and a recognizable goal (research competitors, build a persona, understand churn, find VOC language), skip the questions, state your plan in one sentence, and start Step 1. Only ask when something essential is genuinely missing — product identity or target segment, for example. Don't ask all five questions before doing anything.

Step 1 — Pick sources based on ICP type

Before calling anything, decide which sources are worth hitting for this specific audience:

Minimum viable run: Reddit + one review site. Add supplementary sources only when the minimum doesn't produce enough signal, or when the ICP table above calls for them.

Step 2 — Run targeted scrapes

Pull from at least 2 sources. Single-source findings are low confidence by definition.

Reddit — the highest-signal source for most ICPs:

pythonCopy code
reddit_scrape(
    searches=["[product category] frustrations", "[competitor name] problems"],
    sort="top",
    time="year",
    max_items=50,
    skip_comments=False,
    search_posts=True,
    search_comments=True
)

For specific subreddits, pair with start_urls:

pythonCopy code
reddit_scrape(
    start_urls=["https://www.reddit.com/r/marketing/"],
    searches=["CRM"],
    sort="top",
    time="year",
    max_items=30
)

YouTube comments — rich qualitative data:

pythonCopy code
# Step 1: find the relevant videos
youtube_videos_search_top(query="[product category] honest review", max_results=5, sort_by="views")

# Step 2: mine comments from the top results
youtube_comments_search(
    start_urls=["https://www.youtube.com/watch?v=VIDEO_ID_1", "https://www.youtube.com/watch?v=VIDEO_ID_2"],
    max_comments=100,
    comments_sort_by="0"   # "0" = top comments, "1" = newest
)

X/Twitter — complaints, frustrations, and niche conversations:

pythonCopy code
x_tweets_search(
    search_terms='"[product name]" frustrating OR broken OR switched OR canceled',
    max_items=50,
    min_faves=5
)

Review sites (G2, Capterra, Trustpilot):

pythonCopy code
# G2 reviews for a specific product
web_pages_scrape(
    url="https://www.g2.com/products/[product-slug]/reviews",
    ai_query="Extract the top complaints and pain points from customer reviews. Include verbatim quotes.",
    use_proxy=True
)

TikTok — consumer conversations and trending frustrations:

pythonCopy code
tiktok_videos_scrape(
    search_queries=["[product category] problems", "[competitor name] review"],
    results_per_page=30
)

Google discovery — find threads and communities you haven't thought of:

pythonCopy code
google_search_results_search(
    query='site:reddit.com "[product category]" "I switched" OR "I quit" OR "stopped using"',
    num_results=20
)

Step 3 — Extract signal from raw data

For each source, extract into this structure:

Step 4 — Synthesize across sources

After pulling from 3+ sources, synthesize into the research report format in references/synthesis-templates.md. The report includes:

• Top themes ranked by frequency × intensity
• VOC quote bank organized by theme
• Confidence labels on every finding
• Source bias notes

Step 5 — Build personas (optional)

Only build personas if you have ≥5 independent data points from a consistent segment. If not, say so and describe what additional research is needed first.

Persona template is in references/synthesis-templates.md.

Questions to ask before starting

Only ask what's genuinely missing. If the product and goal are clear, go. If not, lead with these — one or two at a time, not all at once:

1. What's the product? (if not obvious from context — a URL works)
2. What's the goal? Improve messaging? Build personas? Understand churn? Find product gaps?
3. Who is the target segment? (all customers, a specific tier, churned users, prospects who didn't convert)
4. What do you already have? (transcripts, surveys, tickets, nothing)
5. What deliverable do you need? (synthesis report, quote bank, persona, competitive language comparison)

Deliverables

Ask which one(s) the user needs before generating: