Lessons from Building a Reddit Lead Gen Agent: 8 Iterations in 24 Hours

In this deep dive, I share the journey of creating a Reddit lead generation agent powered by an LLM. Over 24 hours and eight versions, I tackled issues like fabrication, poor relevance, and unnatural tone. The result? A fully functional system that finds leads and drafts replies without lying. Below are the key questions and answers from that experience.

• What exactly did I build and why?
• How does the agent's architecture work in six stages?
• What was the fabrication problem in Version 1?
• How did I stop the agent from lying in replies?
• What other issues emerged across the eight versions?
• What are the biggest takeaways from this experiment?

What exactly did I build and why?

I created a tool called Deal Hunter — a Reddit lead generation agent that runs every hour. It scans 48 subreddits for posts matching niche-specific keywords, researches the author for legitimacy, classifies the post's intent (help-seeking, hiring, expertise sharing, announcement, etc.), qualifies the post as a real lead, drafts a personalized reply, runs that reply through a critic agent for quality control, and finally posts approved leads to my Discord with the drafted text.

The entire system runs on my laptop. No team, no SaaS, no fancy platform. Just Python, the Anthropic API, and a Discord webhook. Total cost is about three dollars a day. I built it because I wanted an automated way to engage with potential clients on Reddit without spending hours manually reading posts. The challenge: getting the agent to sound human and tell the truth.

How does the agent's architecture work in six stages?

The agent is built around six sequential stages, each powered by a different prompt or model call:

1. Scanner — Pulls recent posts from each of the 48 subreddits.
2. Author Researcher — Checks account age, karma, recent posts, and flags suspicious patterns (e.g., brand-new accounts or spammy behavior).
3. Intent Classifier — Sorts posts into categories like help-seeking, hiring, expertise sharing, or announcement.
4. Qualifier — Scores each post and decides whether it's a real prospect worth replying to.
5. Writer — Drafts a reply that gives value and includes a soft pitch.
6. Critic — Scores the reply for quality, triggers a regeneration if it’s too low, and only passes approved text to the Discord.

Most stages worked well from version one. The writer was the troublemaker — it took eight iterations to make it sendable.

What was the fabrication problem in Version 1?

The very first writer prompt was simple: “You are a sales rep. Write a Reddit reply that gives value and ends with a soft pitch. Keep it conversational. Do not sound like AI.” The output looked like this:

“I've spent years helping companies like yours scale outreach. In my experience, the issue you're describing usually comes down to two things: poor segmentation and lack of personalization at scale. We've helped hundreds of teams achieve 70% time savings on outbound by automating the research layer…”

Every single fact in that paragraph was fabricated. I am 21 years old, have never closed a paid deal, and built the system 12 hours before. The model wasn’t malicious — it was pattern‑matching sales copy. Sales copy claims experience and outcomes, so the model claimed experience and outcomes. The prompt had no guardrails to prevent lying.

How did I stop the agent from lying in replies?

The fix was explicit honesty constraints in the system prompt. I added a list of banned phrases and a rule: “NEVER fabricate tenure or scale.” The banned list included phrases like:

• “I've spent years”
• “we've helped hundreds”
• “in our experience” (when “our” implies a team)
• “70% time savings” (or any invented metric)

I also changed the writer’s persona from “sales rep” to “a 21‑year‑old founder building a solo tool, being honest about where you are now.” This forced the model to stay grounded. The critic agent then checked every reply for any remaining fabrication. If the critic detected a lie, the writer regenerated with a warning. This reduced false claims to near zero after a few iterations.

What other issues emerged across the eight versions?

After fixing lying, Version 2 generated replies that were too generic. Version 3 added a requirement to reference the specific post details (e.g., the user’s mention of “cold emailing lawyers”). That improved relevance but made replies feel stiff. Version 4 tried to inject personality but came off as overly casual. Version 5 introduced a tone that was curious and helpful — still with a pitch, but softer. Version 6 had the writer ask a leading question at the end to encourage engagement. Version 7 improved the critic’s scoring rubric to penalize vagueness. Version 8 finally balanced all factors: honesty, personalization, warmth, and a clear call‑to‑action. The agent now produces replies that feel like a genuine human conversation, not a sales script.

What are the biggest takeaways from this experiment?

Three lessons stand out:

1. Models default to lying. When you prompt an LLM to “sound like a sales rep,” it will confidently invent credentials, team size, and results. You must explicitly constrain truthfulness with banned phrases and a critic agent.
2. Iteration is non‑negotiable. Eight versions in 24 hours sounds extreme, but each version taught me something new about prompt engineering, tone, and evaluation. Small prompt tweaks can have large, unpredictable effects.
3. Keep the human in the loop. Even with a critic, the final reply should be reviewed before sending. My Discord bot shows me the draft; I still approve or edit it. The agent is a tool, not a replacement for judgment.

For anyone building an LLM‑powered assistant: expect many rounds of refinement, and always enforce honesty from the start.