Templates
Free, production-tested prompt templates.
10 templates we use in real engagements — system prompts, refusal libraries, eval seeds, structured-output schemas. Copy, adapt, ship.
These aren't demo prompts. Each one is a stripped-down version of something running in a client system, which is why they look less clever than the prompts you see in blog posts — production prompts spend most of their length on refusals, escalation criteria, and what to do when the retrieved context is empty. That unglamorous middle section is the part that decides whether the thing survives contact with real users.
How to use these without getting burned
- Replace the voice block, keep the guardrails
- The brand-voice section of every system prompt is meant to be rewritten. The refusal and escalation sections are not — those are the parts tuned against real failures, and they're the first thing people delete because they read as verbose.
- A prompt is not a system
- A support prompt with no retrieval behind it will confidently invent your return policy. Most of these assume something specific underneath — hybrid search, a reranker, a tool allowlist. The playbooks cover what goes under them.
- Start the eval suite before you tune
- The eval starter is 20 cases because 20 is the number people will actually maintain. Without it, prompt iteration is just vibes — you fix one behaviour, silently break another, and find out from a customer.
- System prompt
Customer support agent system prompt
Production-grade system prompt for a tier-1 customer support agent. Citation-required, refusal-aware, escalation-routed.
When to use: Use when building a customer-facing support chatbot over a knowledge base. Pair with hybrid search + reranking. Tune the brand voice block to your team.
Read template - System prompt
Inbound lead qualification system prompt
BANT-style qualification flow that engages inbound leads in 60 seconds and books AEs only when a real fit appears.
When to use: Use for multi-channel inbound qualification (web chat, email reply, SMS). Customize the qualification fields to your ICP.
Read template - Pattern
Refusal pattern library
The refusal patterns we add to every agent. Copy + adapt to your domain — these are the safety net.
When to use: Add these patterns to your agent's system prompt or middleware layer. Each one prevents a class of failure mode.
Read template - System prompt
Citation-required RAG prompt
The citation pattern that drops hallucination 80%+ in production RAG systems. Pair with hybrid search + reranking.
When to use: Wrap your RAG generation step. Numbered citations match retrieved chunks. Refuse if no relevant context.
Read template - System prompt
Code review agent system prompt
The prompt behind our Auto Issue Resolution case study — for an agent that reviews PRs, not writes them.
When to use: Adapt for any code-review augmentation. Used in conjunction with read-only repo access + sandboxed execution.
Read template - Eval
Eval suite starter (20 cases)
The seed structure for the 20-case eval suite we ship with every agent on day one. Replay weekly, grow from production.
When to use: Start here on day one. Add 5 cases per week from production failures. Run on every prompt change in CI.
Read template - Schema
Structured JSON output prompt
How to get reliable JSON out of an LLM without parsing pain. Use sparingly — Claude and GPT both support structured-output APIs natively.
When to use: When you need a downstream system to consume the model's output. Pair with the structured-output API of your model where available.
Read template - System prompt
Voice agent greeting + handoff
The opening 30 seconds of a voice agent — set expectations, qualify the call, route fast.
When to use: Use as the system prompt for Twilio / Vapi / Retell voice agents. Optimize for latency, not eloquence.
Read template - System prompt
Document extraction prompt
Extract structured fields from PDFs, invoices, contracts, or forms — with explicit confidence scoring per field.
When to use: Pair with a vision model or OCR pipeline. Validate every field against business rules before posting downstream.
Read template - Pattern
Approval-gated action prompt
The pattern that makes agent writes safe — draft an action, request approval, execute only on confirmation.
When to use: Wrap any agent action that has irreversible consequences (sending email, creating tickets, posting to ledger).
Read template
