Most prompts fail for the same reason: they describe what someone wants in a paragraph and hope the model figures out the rest. Production agents don't run on hope. This is the structure we use at PlatOS to write system prompts that hold up under real traffic, real edge cases, and real clients.

The five-part system prompt

Every reliable agent prompt we ship is built from the same five blocks, in this order. Skip one and the cracks show up in production.

1. Role and objective

State one job, sharply. An agent with one clear objective outperforms a "do-everything" assistant every time. If you can't write the role in a single sentence, the agent is doing too much — split it.

You are a scheduling agent for a home-services business.
Your only job is to book qualified leads onto the calendar.

2. Context

Tell the agent who it's talking to and what data it can see. The model can't infer your business — spell out the user, the situation, and the inputs available on each turn.

3. Tools

List each tool and, more importantly, when to call it and when not to. Models over-call tools when the boundaries are fuzzy. Be explicit: "Call check_availability before offering any time slot. Never invent a slot."

4. Guardrails

Define what the agent must never do, and where it escalates. This is the block that keeps you out of trouble. Cover the off-ramp for anything outside scope — a human handoff, a canned response, a refusal — so the agent never improvises its way into a problem.

5. Output format

Show, don't describe. One concrete example of the exact output you want beats three sentences describing it. If structure matters, give a filled-in sample, not a schema in prose.

Techniques that pull weight

Positive and negative examples. Show a good response and a bad one. The contrast teaches faster than either alone.
Step-by-step reasoning for multi-step tasks. When the job has sequence or logic, tell the model to work through it in order before answering.
Explicit tags for structured output. Wrap structured sections in clear delimiters (XML-style tags work well) so downstream code can parse them reliably.
Tight scope beats clever wording. A narrow, well-bounded agent is more valuable than a clever one that tries to do everything. Constraint is a feature.

A reusable skeleton

Start every new agent from this and fill the blanks. It's deliberately boring — boring is what survives production.

# Role
You are [one job, stated sharply].

# Context
The user is [who]. On each turn you can see [what data].

# Tools
- [tool]: call when [condition]. Never call when [condition].

# Guardrails
- Never [thing].
- If [out-of-scope situation], then [escalation path].

# Output
Respond in this format:
[concrete filled-in example]

How to ship it

Write the prompt, then break it on purpose. Feed it the weird inputs — the angry customer, the half-filled form, the question it shouldn't answer. Watch where it improvises, then tighten the guardrail that let it. A prompt isn't done when it works on the happy path; it's done when it fails safely on the ugly ones.

The goal isn't a prompt that sounds smart. It's an agent that does one job, the same way, every single time.

The Prompting Playbook