Building a Lead-Gen Pipeline with Liquid AI

Last month I needed to send 500 personalized cold emails to niche SaaS founders, but my manual copy‑paste process ate two days and still left half the list untouched.

By the end of this article you’ll have a working Liquid AI flow that scrapes LinkedIn, writes personalized copy, and pushes leads to your CRM—plus you’ll know where it tends to break.

Scenario: The Cold‑Email Bottleneck I Hit

I was running a solo consultancy and needed to reach founders of micro‑SaaS tools that sell for under $5k/mo. My list came from a LinkedIn Sales Navigator export, but turning those profiles into custom first lines took forever. I tried a generic template and got a 0.8% reply rate—basically noise.

I decided to automate the research and copy steps with Liquid AI, hoping to keep the human touch while cutting the time from two days to a few hours. The first attempt failed because the scraper hit a hidden rate limit after about 120 profiles.

Concrete gripe: the Liquid AI API docs said the starter tier allowed 200 requests per minute, but after I burned through 120 calls the response switched to 429 with no warning. I had to add a sleep timer and re‑work the batch logic, which ate three hours of debugging.

Concrete love: the built‑in tone checker flagged a draft that sounded too pushy, letting me swap a line before I sent anything. That tiny guardrail kept my reply rate from dropping into spam territory.

After I added the throttling fix, the flow ran smooth for the full 500‑record batch. The whole process—scrape, enrich, write, push—took about 80 minutes on a modest laptop.

Why does the LinkedIn scraper stall after 100 profiles?

The stall isn’t a bug in Liquid AI; it’s a side effect of how LinkedIn throttles external tools that scrape public profiles. When you pull more than ~100 pages in a short window, LinkedIn serves a CAPTCHA or returns empty JSON. Most guides miss this because they test with tiny lists.

What works is to mimic a human pace: request one profile, wait 12‑15 seconds, then request the next. I wrapped the scraper in a simple loop that respects that delay. The trade‑off is slower throughput, but you avoid getting blocked entirely.

If you try to push harder by using multiple IP addresses or rotating user‑agents, you’ll likely trigger a temporary ban that lasts 24‑48 hours. I learned that the hard way after a night of failed runs.

What most guides get wrong about prompt tuning

Many tutorials tell you to drop a long prompt into the model and hope for the best. They suggest something like “Write a personalized cold email based on the founder’s background.” The output is often bland or repeats the same phrasing.

What actually moves the needle is to give the model a clear structure and a few concrete examples. I use a three‑part prompt: (1) role definition, (2) context block with the prospect’s data, (3) a short example of the tone I want.

Here’s the exact prompt I use (copied from my Liquid AI workflow):

You are a friendly SaaS advisor who helps founders grow revenue.
Prospect data: {{first_name}} works at {{company}} as {{title}}. Their recent post: {{linkedin_post}}.
Write a 120‑word cold email that starts with a specific compliment about their post, mentions one way your service could help them hit a $10k/mo goal, and ends with a low‑pressure call to action to book a 15‑min chat. Keep the tone helpful, not salesy.
Example: Hi Alex, I loved your recent post about reducing churn through in‑app messaging. At {{company}} we’ve helped similar teams lift activation by 22% using targeted onboarding flows. Would you be open to a quick 15‑min call next week to see if we can do the same for you? Best, Sam

The double curly braces are Liquid AI’s placeholder syntax; they get swapped with scraped fields before the model runs. Supplying the example line dramatically cuts the variance in output.

How to debug when the AI copy sounds robotic

Even with a good prompt, sometimes the generated email feels like a template. The first thing I check is the temperature setting. Liquid AI defaults to 0.7, which is fine for creativity but can drift into repetition if the context block is too short.

I lower the temperature to 0.5 and add a “variability instruction”: “Avoid starting more than two sentences with the same word.” That small tweak forces the model to mix up openings.

If the copy still feels off, I look at the scraped data. Missing or malformed fields (like a blank {{linkedin_post}}) cause the model to fall back on generic phrasing. I added a validation step that skips any prospect where the post field is empty and logs the ID for manual review later.

Finally, I run a quick spam‑score check using a free tool like Mail‑Tester. If the score spikes above 5, I revisit the tone checker setting in Liquid AI and tighten the aggression threshold.

Concrete example: a Liquid AI prompt that actually works

Below is the full snippet I paste into a Liquid AI “Text Generation” node. It pulls data from a previous “HTTP Request” node that fetches a LinkedIn profile via a public scraping API (I use Apify’s LinkedIn scraper, which costs $0.005 per profile).

# Node: Generate Email
prompt: |
You are a friendly SaaS advisor who helps founders grow revenue.
Prospect data: {{first_name}} works at {{company}} as {{title}}. Their recent post: {{linkedin_post}}.
Write a 120‑word cold email that starts with a specific compliment about their post, mentions one way your service could help them hit a $10k/mo goal, and ends with a low‑pressure call to action to book a 15‑min chat. Keep the tone helpful, not salesy.
Example: Hi Alex, I loved your recent post about reducing churn through in‑app messaging. At {{company}} we’ve helped similar teams lift activation by 22% using targeted onboarding flows. Would you be open to a quick 15‑min call next week to see if we can do the same for you? Best, Sam
model: liquidai/text‑fast
temperature: 0.5
max_tokens: 250
stop: ["\n\n"]

The node returns the raw email text, which I then pipe into a “Gmail Send” node (via Zapier automations) after a quick manual glance. The cost per email breaks down to roughly $0.008 for the scrape, $0.012 for the AI generation, and $0.00 for the send—about two cents per outreach.

Pricing take: is the $39/mo plan worth it?

Liquid AI’s starter plan is $39 per month and includes 50,000 token generations, which translates to roughly 2,500 emails at the settings I use. For a solo operator sending 500‑1,000 emails a month, that’s more than enough.

I think $39/mo is fair for the time saved—my manual process would have cost me at least eight hours of labor, which at a $50/hr rate is $400. The AI cuts that to under two hours, so the plan pays for itself in less than a week of use.

If you only need occasional bursts, the free tier (10,000 tokens) might cover a test run, but you’ll hit the limit fast once you go beyond a couple hundred prospects.

Putting it all together: the flow diagram (in words)

1. **Scrape** – HTTP request to Apify LinkedIn scraper, returns an array of profiles.
2. **Enrich** – Loop through each profile, pull {{first_name}}, {{company}}, {{title}}, {{linkedin_post}}.
3. **Validate** – Skip any record where {{linkedin_post}} is empty; log the ID.
4. **Generate** – Feed the data into the Liquid AI Text Generation node with the prompt shown above.
5. **Review** – Optional manual step: view the first 10 outputs in a Google Sheet.
6. **Send** – Pass the approved email to a Gmail Send node (via Zapier) with the prospect’s email address.
7. **Track** – Log sent date and message ID to Airtable for follow‑up.

Each step is a separate node in Liquid AI’s visual builder; you can copy‑paste the JSON export I’ve attached in the /vault blueprint to get a working copy in under five minutes.

If you want the deep cut on this, deeper coverage of AI agent platforms.

If you’d rather skip the build and deploy a working version in an afternoon, we’ve packaged this workflow as a blueprint at deepusecase.com/vault.