Appearance
Guidelines
Guidelines are organizational policies that shape how agents behave — from writing tone to data handling to decision-making boundaries.
Where guidelines live
Guidelines are managed inside Knowledge (/knowledge) alongside documents and context files — they are no longer a separate top-level surface. Everything below still applies; only the location changed.
Guidelines are your guardrails. They let you encode company policies, compliance rules, and style preferences so that every agent follows them consistently. Instead of repeating "never share customer emails" in every workflow, you write it once as a guideline and it applies everywhere.
What It Does
- Enforces consistency — Every agent in your workspace follows the same rules, every time.
- Protects sensitive data — Data guidelines prevent agents from exposing information they should not.
- Sets boundaries — Decision guidelines define what agents can decide on their own and what requires escalation.
- Controls tone — Tone guidelines ensure all agent-generated content matches your organization's voice.
Guideline Categories
| Category | Purpose | Example |
|---|---|---|
| Tone | Writing style and voice | "Use formal language for external communications. Keep internal messages conversational." |
| Data | How to handle sensitive information | "Never include raw customer emails in reports. Redact Social Security numbers." |
| Decision | Boundaries for autonomous choices | "Escalate any expenditure over $10,000 to a manager for approval." |
| Safety | Hard limits on agent behavior | "Never delete production data. Never send communications to customers without approval." |
Tone Guidelines
Tone guidelines shape how agents write. They apply to every piece of text an agent generates — emails, reports, summaries, chat responses. You can be as specific as you need.
Examples:
- "Write in active voice. Avoid passive constructions."
- "Use bullet points for lists of three or more items."
- "Address clients by first name. Sign off with the team name, not an individual."
Data Guidelines
Data guidelines control how agents handle sensitive information. They are especially important for workflows that touch customer records, financial data, or personal information.
Examples:
- "Aggregate customer data to the regional level. Never include individual records."
- "Mask credit card numbers, showing only the last four digits."
- "Do not store API keys or passwords in artifacts."
Decision Guidelines
Decision guidelines tell agents where their authority ends and human judgment begins. They prevent agents from making high-stakes decisions without oversight.
Examples:
- "If a refund request exceeds $500, escalate to the support manager."
- "Do not change pricing without approval from the pricing team."
- "When two data sources conflict, flag the discrepancy instead of choosing one."
Safety Guidelines
Safety guidelines are hard stops. They define actions the agent must never take, regardless of context. These are your non-negotiable rules.
Examples:
- "Never execute DELETE operations on production databases."
- "Never send external communications without a human reviewing the final draft."
- "Never share internal financial projections outside the organization."
WARNING
Safety guidelines should be written as clear, absolute rules. Avoid ambiguous language like "try to avoid" or "generally do not." Use "never" and "always" so the agent has no room for misinterpretation.
Key Properties
| Property | Description | Options |
|---|---|---|
| Scope | Who the guideline applies to | workspace (all agents), workflow (a specific workflow), user |
| Activation Mode | When the guideline is active | always (every run) or when_relevant (semantic match) |
| Active | Whether the guideline is enforced | true / false |
| Applies To | Specific workflows this guideline targets | Workflow IDs (only when scope is playbook) |
Understanding Scope
| Scope | What it means | Use it for |
|---|---|---|
| Workspace | Applies to every agent in every workflow across your entire workspace | Company-wide policies, universal tone, safety rules |
| Workflow | Applies only when a specific workflow runs | Workflow-specific instructions, workflow-level overrides |
| User | Applies to interactions with a specific user | Personal preferences, individual style settings |
TIP
Start with workspace-level guidelines for your most important policies. Add workflow-level guidelines only when a specific workflow needs different rules.
Understanding Activation Modes
Activation mode controls when a guideline gets included in the agent's instructions.
Always
Guidelines set to always are included in the agent's prompt on every single run and every chat message. Use this for rules that must never be forgotten:
- Safety rules ("Never delete production data")
- Universal tone requirements ("Use formal language")
- Core data policies ("Redact personal information")
When Relevant
Guidelines set to when_relevant are only included when the agent's current task is semantically related to the guideline's content. Pencel uses semantic matching to determine relevance automatically.
This is useful for guidelines that only matter in certain contexts:
- "When writing quarterly reports, include a comparison to the previous quarter" — only relevant during reporting tasks
- "When responding to customer complaints, acknowledge the issue before offering solutions" — only relevant in customer-facing workflows
INFO
when_relevant guidelines keep the agent's instructions lean. Instead of loading every guideline every time, the agent only sees the ones that matter for the task at hand. This improves focus and reduces noise.
How Many Guidelines Should You Have?
There is no hard limit, but here are some practical recommendations:
| Guideline count | Guidance |
|---|---|
| 5-10 | Good starting point for most teams |
| 10-25 | Typical for mature workspaces with multiple workflows |
| 25+ | Consider using when_relevant activation mode so agents are not overwhelmed |
Too many always guidelines can dilute their impact. If you find yourself writing dozens of guidelines, switch most of them to when_relevant and reserve always for your most critical rules.