Maximum Character Length in Guidance
Make sure that you are optimizing your agent for the job they were “hired” for. Do not prompt-bomb the guidance and write a Da Vinci Code length guidance prompt. It’s hard for a human to read and hard for AI to understand.
So, what is the ideal length of your guidance? First, factor in that ZVA is not only digesting your guidance prompt but also,
- Length of the conversation - which we are optimizing using summary.
- Number of skills and their complexity.
- How tool responses are being returned.
- How guidance is written as well (the architecture and formatting matter).
Best Practice:
Target for one agent’s guidance prompt:
- 2,000–4,000 characters
- (~500–1,000 tokens)
Why this range works best
- Enough room for:
- role definition
- goals
- tone
- guardrails
- tool-calling rules
- Short enough that the model consistently attends to it
- Easier to debug when behavior goes sideways
Upper limit: 8,000 characters. You won’t be stopped, but you also won’t see the results you nor your users want.
Once it goes past this, you’re likely to have repeating or conflicting prompts, poor structure and honestly, should be using another agent for that subject matter.
You do not want to use one agent for all use cases… that’s usually how users get into trouble and the result is a over-engineered prompt. Best practice is to do agent-to-agent transfer or routing based on what the user needs before reaching the agent. You can “transfer” to another agent without the user knowing, in order to keep things simple and structured. Just like you would hire one human for a specific job role and not have your CEO also be a Loan Officer, HR Benefits Specialist, Payroll Manager and Engineering manager.
Hope that helps!