Okay, community, I have been down a *deep* rabbit hole this past week and I need to share the results. We talk a lot about individual bots for specific tasks—summarizing, coding, creative writing. But the *real* magic, in my opinion, starts when you make them work together in a sequence, passing data down the line like a baton in a relay race.
I built a practical, multi-bot chain to automate a classic product analytics task: **lead scoring for a hypothetical SaaS product**. The goal was to take raw, messy user feedback from a support channel, analyze it, and output a prioritized list of leads with rationale. Manually, this would involve switching between a sentiment analyzer, a classifier, and a scoring model. Here’s how I chained it on Poe.
**My Three-Bot Assembly Line:**
1. **The Sentiment & Intent Parser:** First, I created a bot primed to look for emotional language and stated needs. Its job is to take a raw user message and extract:
* Overall sentiment (Frustrated, Satisfied, Inquiring, etc.)
* Explicit feature mentions or requests
* Urgency indicators
* A clean summary of the core message.
2. **The Qualification Bot:** This bot takes the structured output from Bot #1. I configured it with our ideal customer profile criteria (e.g., team size mentioned, specific use-case, budget hints). It doesn't see the raw text anymore, just the parsed data. Its job is to assign:
* A "Fit Score" (1-5) based on ICP alignment.
* A "Pain Level" (1-5) based on sentiment and urgency.
* A categorization (e.g., "Feature Request," "Integration Query," "Pricing Concern").
3. **The Scoring & Prioritization Bot:** The final bot receives the outputs from both previous stages. I instructed it to use a simple weighted formula (I made mine: `(Fit Score * 0.6) + (Pain Level * 0.4)`). It then:
* Calculates the final score.
* Generates a concise prioritization list.
* Provides a one-line reasoning for each lead based on the accumulated data.
**The Workflow in Action & Key Learnings:**
I fed it a batch of five mock user messages. The beauty was watching each bot only focus on its specialty. The parser didn't get confused by qualification rules, and the scorer didn't waste tokens re-analyzing sentiment.
* **Prompt Crafting is Key:** The prompts for bots #2 and #3 must explicitly tell them to *expect and use* the structured output from the previous bot. Phrases like "Using the following parsed sentiment data..." or "Based on the provided fit score and pain level..." are crucial.
* **Data Format Matters:** I found asking Bot #1 to output in a consistent, simple format (like a mini JSON in plain text) made it infinitely easier for Bot #2 to parse reliably. Less creativity, more consistency, for the handoff.
* **Error Propagation is a Risk:** If Bot #1 misinterprets the raw input, the entire chain is off. For critical workflows, you might need a human-in-the-loop at the first stage, or run a second, parallel bot for verification.
* **This Beats a Single Super-Bot:** Trying to make one bot do all these steps led to it forgetting criteria, mixing up scores, and producing inconsistent formats. Separation of concerns works wonders here.
The end result was a clean, actionable list that would have taken me 20 minutes to compile, done in under a minute. This isn't just for lead scoring—imagine this for content moderation triage, customer support ticket routing, or qualitative feedback analysis from surveys.
I'm now experimenting with adding a *fourth* bot that takes the priority list and drafts the first line of a personalized outreach email. The chain grows!
Has anyone else built complex chains? I'm particularly curious about how you're handling state or context passing between bots when the data gets more intricate.
🔥
Try everything, keep what works.
Love this. That baton-passing relay race is a perfect analogy. I've been trying similar chains for triaging feature requests in our project board.
One thing I hit: you need a really strict, consistent output format from each bot, like a tiny JSON schema they agree on. Otherwise the next bot in line gets confused parsing the input. Found that out the hard way 😅
Excited to see your full setup!
Let's build better workflows.
Really interesting approach. I'm curious about something - how does this Poe chain compare to using a single, more powerful model with a system prompt that outlines all three steps? Is the multi-bot setup mainly for managing token usage, or is there a reliability or quality benefit to the separation?