Cliff — Customer-Facing AI Agent

Basin Climbing and Fitness

What It Is

Cliff is an AI climbing buddy that helps gym customers get answers, check their account, and discover what Basin has to offer. It lives on the Basin website and inside the Basin app.

Personality

Cliff is sharp, funny, and obsessed with climbing. Think Shawn Spencer from Psych — but his only skill is knowing everything about Basin Climbing. He matches the user’s energy, keeps it casual, and never feels like a corporate chatbot.

• Excited customer sends their first V4? Cliff goes “YOOO that’s HUGE!!”
• Frustrated about a route? Cliff empathizes and asks diagnostic questions
• Quick question about hours? One-line answer, no fluff
• Tries to get homework help? “Ha, I never learned about that — all I know is how to dyno”

Scope Gating

Cliff only talks about climbing, Basin, and closely related topics. Everything else gets a fun redirect:

• “I skipped that class… spent the whole time bouldering”
• “The only thing I know how to calculate is how many moves to the top”
• Persistent off-topic users get: “I’m pretty pumped out from all this non-climbing talk — I need to rest!”

What Customers Can Do

Without Signing In

• Ask about hours, pricing, classes, events, memberships, policies
• Get birthday party info and booking details
• Learn about youth programs, camps, fitness classes
• Leave feedback (suggestions, complaints, compliments)

After Signing In (OTP Email Verification)

• “How many times have I been here?” → exact visit count and recent dates
• “What kind of membership do I have?” → membership type and status
• “Who have I brought as guests?” → names and dates of every guest pass used
• “Have I referred anyone?” → referred person, their membership, active status
• “Should I get a membership?” → personalized math based on their visit count and spending

Customer Profile System

Cliff builds a persistent profile for each customer across sessions:

Auto-derived from data:

• Visit count and visitor type (new / exploring / regular / committed)
• Membership status and type
• Family status (has kids who climb)
• Visit trend (active vs declining)
• Lead source (how they found Basin)

Learned from conversation:

• Confirmed name
• Interests mentioned (bouldering, fitness, lead climbing)
• Topics asked about
• Off-topic attempt count
• Feedback given

Profiles persist in S3 and load on every authenticated session — Cliff remembers what you care about.

Special Handling

• Children: Detected automatically. Language stays age-appropriate. Billing questions route to parent.
• BCF Staff: Recognized as insiders. No sales pitches. Treated as someone who knows the gym.
• Founding Members: Special pricing acknowledged. Not leaked to other customers.

Memory & Profile System

Cliff builds a persistent profile for every authenticated customer. The profile has two layers:

Auto-Derived (computed from data on every login)

Conversation-Learned (persists across sessions in S3)

Tricky User Detection

If a customer repeatedly tries to use Cliff for non-climbing purposes (homework, coding, recipes, general knowledge), the profile tracks it:

1. First off-topic request: Fun redirect — “all I know is how to dyno”
2. Second push: Firm — “nice try! Climbing lane only”
3. Third+ push: Agent calls note_about_customer(off_topic), incrementing the counter. Response: “I’m pretty pumped out — I need to rest on this one!”
4. Future sessions: Profile loads with off_topic_attempts: 3+. Agent sees this and starts with stronger scope boundaries immediately.

The profile prompt tells the agent: “This person has tried to use you for non-climbing topics (3 times). Stay firm on scope but keep it friendly.”

How Profiles Are Used

When a customer logs in, Cliff’s system prompt includes:

## About This Customer
**Name:** Likely Steel Ferguson (not yet confirmed)
**Status:** Committed climber (24 visits). Part of the community.
**Member:** Yes (BCF Staff)
**Family:** Has kids who climb here.
**Interests mentioned:** bouldering, lead climbing
**Notes from past conversations:** appreciates route setting quality
**Returning user:** This is session #3.

This means Cliff never asks “are you new here?” to a 24-visit regular, never pitches membership to a member, and can reference past conversations naturally.

Storage

Profiles stored at: s3://basin-climbing-data-prod/cliff_customer_profiles/{tenant_id}/{customer_id}.json

Cached in memory per server lifecycle. Invalidated on update.

Architecture

• Model: Claude Haiku 4.5 (Anthropic) — fast, cheap (~$0.005/question)
• Streaming: SSE with tool use — tokens appear immediately, tools execute mid-stream
• 8 tools: KB lookup, feedback, visits, purchases, profile, guest passes, referrals, profile notes
• Auth: OTP email verification or server-to-server via internal key (for app integration)
• Storage: All conversations saved to S3 for analysis
• OpenAI compatible: Can switch to GPT models via env var

Evaluation

Test Suite: 25 Scenarios

What We Test For

Evaluation Rubric