An agent with a good system prompt does decent work. An agent with a playbook does it the way you want, every time.
Playbooks are structured instruction sets you attach to agents. They define what the agent should handle, how it should think, what it should never do, and how it should communicate. Think of them as standard operating procedures for AI.
Why not just use the system prompt?
You can. But system prompts are free-form text, and they tend to get long and inconsistent. Playbooks give you five structured sections that force clarity:
Intent definitions. What situations should this agent handle? A research playbook might say: "Deep-dive analysis on companies, industries, or technologies. Not quick lookups or simple fact checks."
Decision hierarchy. How should the agent prioritize when things conflict? "Accuracy over speed. Primary sources over secondary. Recent data over historical."
Business guardrails. Hard boundaries the agent must never cross. "Never fabricate data points. Never present estimates as facts. Always cite sources."
Communication rules. Tone, format, structure. "Use headers and bullet points. No marketing language. Include confidence levels on claims."
Additional context. Domain knowledge, examples, reference materials. "Our target market is mid-market SaaS. Competitors include X, Y, Z."
Creating a playbook
Go to Playbooks and click Add Playbook. Pick from four templates (research, writing, code review, customer support), start from scratch, or let AI generate one from a description.
Templates pre-fill the five sections with sensible defaults. The research template, for example, comes with source verification rules, structured output formats, and depth-over-breadth priorities. Edit what doesn't fit, keep what does.
Attaching playbooks to agents
In your agent's settings, under Advanced Settings, you'll find a Playbooks picker. Select one or more playbooks. The agent receives the playbook instructions alongside its system prompt and task.
Multiple playbooks compose. A research agent might get both a "research methodology" playbook and a "company analysis" playbook. The agent follows both.
With vs. without
Run the same research task twice: once without a playbook, once with a research playbook that specifies structured output, source citation, and confidence levels.
Without: you get a wall of text with decent information but inconsistent formatting and no sources.
With: you get a structured report with headers, cited sources, confidence ratings, and a consistent format you can rely on.
The difference compounds. One run, it's nice to have. Twenty runs over a month, it's the difference between useful output and output you actually use.
Playbooks in workflows
Playbooks work especially well in workflows. Each workflow step runs its assigned agent with that agent's playbooks — so step 1's research agent follows its research methodology, step 2's analysis agent follows its own rules, and so on. The workflow coordinates the handoff.
Try it
Go to Playbooks, create one for your most common task type, and attach it to an agent. Run the agent with and without. The difference is immediate.