Every agent in Rival has three fields: a name, a description, and instructions. Users see the first two. The instructions are what actually shape how your agent behaves, and they are usually the difference between a reliable agent and an unpredictable one.
When an agent underperforms, most builders add more tools. (Tools are the actions an agent can take, like sending an email or searching the web.) But the fix is almost always clearer instructions. Here are the five parts that get you there.
Write Instructions in Quick Five Steps
Break your agent's instructions into five labeled sections. Each one answers a question your agent would otherwise have to guess at.
1. Identity: Who is the agent?
Give your agent a role, not just a task. A role gives it a consistent point of view to work from, so it responds the same way across different requests.
Avoid: “You are a helpful assistant.”
Instead: “You are an outreach specialist who writes concise cold emails for university organizations.”
2. Objective: What is it trying to achieve?
State the outcome you actually want in a single line, so the agent knows what success looks like.
Example: “Write short outreach emails that get replies while maintaining a professional tone.”
3. Workflow: What steps should it follow?
Lay out the process in order so the agent works the same way every time, guaranteed.
Example:
Confirm the goal of the email and the response you want.
Gather details about the club and the recipient. If anything important is missing, ask before drafting.
Draft a short, personalized message.
Review it for clarity, tone, and length.
4. Guardrails: What should it never do?
This is the part most builders skip, and it is exactly where errors sneak in. Spell out the limits so the agent does not improvise when a prompt is vague or unexpected.
Example:
Do not invent facts about the club or the recipient.
Do not imply a relationship or past contact that does not exist.
Do not make commitments or promises on the club's behalf.
Keep claims to what only you know of the club.
5. Context: What background does it need?
Give it the details that shape the output, so you are not correcting the same things on every run.
Example:
Audience: university clubs and organizations
Tone: professional but approachable
Length: under 150 words
Put together, those five sections leave your agent far less room to improvise, and improvising is what makes outputs inconsistent.
Be Specific: It Removes the Guesswork
When instructions are vague, an agent fills the gaps on its own, and you get a different result every time. Specific instructions close those gaps.
Vague: “Summarize documents for users.”
Specific: “Summarize the uploaded document into three sections: key findings, risks, and recommended actions. Keep each section under five bullets, and use only information found in the source.”
The second version tells the agent exactly what “summarize” means here. When an agent seems to be underperforming, it is usually not broken. It is guessing, and your job is to give it less to guess about.
Refine as You Go
Instructions are not a one-time setup. The best builders treat them as a living document and tighten them after each test run. Once your agent runs, look for the moments it possibly went off-track:
Where did it make an assumption it should not have?
Where did it ignore your formatting?
Where did it add information no one asked for?
Each of those is usually a small gap in the instructions, not a flaw in the agent. Fix the wording, run it again, and the output gets steadier.
TL;DR
A great agent usually comes down to great instructions, not more tools. Before you publish, make sure your instructions cover five things:
Identity: who the agent is
Objective: what it is trying to achieve
Workflow: the steps it follows
Guardrails: what it must never do
Context: the background it needs
The more specific you are, the less your agent has to guess. In Rival, clearer instructions almost always mean a better agent.