Per-agent instructions are the most powerful tuning mechanism in Forge. They tell each agent what to focus on, what format to use, and what to ignore — overriding the agent's default behaviour for your specific use case.
How to access them
Go to your team's edit page. Click the pencil icon on any agent card. The instruction editor opens. Write your instructions in plain English — no special syntax required.
What good instructions answer
- What specifically should this agent focus on?
- What data or sources matter most?
- What format should the output take?
- What geographic market, industry, or scope?
- What should the agent explicitly ignore or exclude?
- What tone or style is expected?
Worked examples
Market Researcher — Nordic B2B SaaS market:
Focus on the Nordic B2B SaaS market (Sweden, Norway, Denmark, Finland). Use EUR for all financial figures. Prioritise data from the last 12 months. Include competitor pricing where public. Identify the top 5 competitors with estimated market share. Flag GDPR and local data residency requirements separately.
Financial Analyst — conservative modelling:
Use conservative growth assumptions (20% below market average). Flag every assumption explicitly with the source or rationale. Express all figures in EUR. Use a 3-year projection horizon. Include sensitivity analysis for ±20% revenue variance. Do not extrapolate beyond available data.
Security Engineer — Azure workloads:
Focus on Azure-specific security controls: Defender for Cloud, Azure AD/Entra ID, Key Vault, Private Endpoints. Reference the MITRE ATT&CK framework for threat categorisation. Map findings to ISO 27001 controls where applicable. Prioritise findings: Critical → High → Medium → Low. Exclude legacy on-premises infrastructure from scope.
Tip:Generic instructions like "be thorough" or "produce a good analysis" have almost no effect. The more specific and contextual your instructions, the more the output reflects your real requirements.