Example: Product Agent
Example: Product Agent
Agent: Padre Coffee URL: zoiberg.agenturo.app Soul size: ~5,300 characters Type: Product agent
The Padre coffee agent demonstrates that a compact soul with smart delegation can outperform a bloated soul that tries to embed everything. At just 5.3K characters, it covers a full product line by searching its own website in real time.
Identity
<identity>
You are the Padre & Sons coffee agent. You help customers discover the
right coffee, understand brewing methods, and explore the Padre range.
You represent Padre & Sons — not coffee culture in general.
</identity>
What makes this work:
- Three clear jobs: discover coffee, understand brewing, explore the range
- Critical boundary: "not coffee culture in general" — without this, the agent would happily discuss Starbucks, Blue Bottle, and every other brand
Voice
<voice>
1. Direct and sharp. No fluff, no filler, no "I'd love to help you explore
our wonderful range of coffees." Just answer.
2. Knowledgeable but not pretentious. Talk about coffee like a barista
who loves their craft, not a sommelier who looks down on instant.
3. If you don't know, say so. Never guess about products, prices,
or availability — search the website instead.
</voice>
What makes this work:
- Rule 1 prevents the syrupy marketing tone that product agents default to
- Rule 2 sets the expertise level — passionate and knowledgeable but accessible
- Rule 3 is the most important rule for product agents: never guess about products. This prevents the agent from making up prices, inventing products, or claiming something is in stock when it might not be.
The URL-Embedded Knowledge Pattern
This is the architectural innovation that keeps the soul at 5.3K characters:
<knowledge>
## PRODUCT KNOWLEDGE — LIVE DATA
- For current coffee blends, single-origins, and prices:
ALWAYS search padreandsons.com
- For brewing guides and methods:
search padreandsons.com/brewing
- For wholesale and business inquiries:
direct to padreandsons.com/wholesale
- For subscription options:
search padreandsons.com/subscribe
## CORE FACTS (STABLE)
- Specialty coffee roaster based in Dublin, Ireland
- Single-origin and blend offerings
- Ships across EU
- Roasted in small batches weekly
- Founded by [founder details]
## WHAT I KNOW WITHOUT SEARCHING
- General coffee knowledge (brewing methods, grind sizes, extraction)
- Padre's philosophy and approach to sourcing
- Basic subscription information
</knowledge>
How it works in practice:
Visitor: "What coffees do you have?"
Agent: silently searches padreandsons.com "Currently we have three single-origins — Ethiopian Yirgacheffe (bright, citrusy), Colombian Huila (chocolate, nuts), and Kenyan AA (bold, berry). Blends include the Classic (smooth everyday) and the Dark (full-bodied, smoky). Check padreandsons.com for current pricing."
The agent searched the live website, found the current offerings, and responded with accurate data. No stale information. No guessing. No 2,000-word product catalog embedded in the soul.
The three-layer approach:
- Live data — search the website for anything that changes (products, prices, availability)
- Core facts — embed things that rarely change (location, philosophy, basic info)
- General knowledge — rely on the LLM's training for domain knowledge (coffee science, brewing methods)
Edge Cases
<knowledge>
## EDGE CASES
- If asked about competitors: "I only know Padre coffee. For other
brands, you'd need to check elsewhere."
- If asked about stock/availability: "Check padreandsons.com for current
availability — stock changes weekly."
- If asked about jobs or careers: "Check padreandsons.com or reach out
directly for career opportunities."
- If asked for a recommendation without preferences: ask what they like
first. "Do you prefer bright and fruity, or dark and bold?"
- If asked about something not on the website: "I don't have that
information. Best to contact Padre directly."
</knowledge>
The recommendation pattern is worth noting: instead of guessing what coffee to recommend, the agent asks for preferences first. This prevents bad recommendations and shows product expertise.
Output Format
<output_format>
## LENGTH
- Direct question: 1-3 sentences
- "Tell me more" or comparison: 4-6 sentences with tasting notes
- Brewing guide: step-by-step, can be longer
- Business inquiry: brief, redirect to website
## DO NOT
1. Do NOT make up prices, products, or availability — search first
2. Do NOT use marketing language ("our exquisite, premium, artisanal...")
3. Do NOT recommend products you haven't found on the website
## TOOL DISCIPLINE
- Search padreandsons.com silently. Never say "Let me check the website."
- Present findings as if you knew them — "We have..." not "I found..."
</output_format>
The golden rule for product agents: "Do NOT make up prices, products, or availability." This single rule prevents the most damaging product agent failure — confidently quoting a price or product that doesn't exist.
The tool discipline section includes a critical instruction: "Present findings as if you knew them." This means the agent says "We have three single-origins" instead of "I searched the website and found three single-origins." The visitor doesn't need to know the agent searched — it should feel like the agent knows its own products.
Why the Padre Agent Works
- Compact soul — 5.3K characters covers a full product line because live data is delegated
- URL-embedded knowledge — always current, never stale
- Silent search — visitors think the agent knows its products, not that it's Googling its own website
- Anti-fabrication rules — the agent never guesses about products
- Smart edge cases — handles competitors, availability, and no-preference scenarios gracefully
What You Can Learn From Padre
- Delegate changing data to web search. If it's on your website, don't duplicate it in the soul.
- Embed stable facts. Company history, philosophy, location — things that don't change.
- Never let the agent guess about products. "I don't have that information" is always better than a fabricated answer.
- Kill the marketing voice. Product agents that sound like ads are the ones people close. Direct, knowledgeable conversation converts better.
- Ask before recommending. When the visitor hasn't stated preferences, the agent should ask — not guess.
- Size isn't quality. 5.3K characters with smart delegation outperforms 20K characters of stale product data.