Skills & Rules

import { Aside } from ‘@astrojs/starlight/components’;

Skills and Rules are planned for Phase 5c of Nova's roadmap. They are not yet implemented. This page describes the intended design.

Skills and Rules make Nova’s agents configurable without code changes. Skills are reusable prompt templates shared across agents and pods. Rules are declarative behavior constraints that complement the Guardrail Agent with user-defined policies and pre-execution enforcement.

Skills — Reusable prompt templates

Skills are reusable blocks of prompt text that can be injected into agent system prompts. Instead of duplicating instructions across multiple agents, you define a skill once and assign it to any combination of agents and pods.

Concepts

Concept	Description
Content	The prompt text, optionally with {{param}} placeholders for parameterization
Scope	global (all agents), pod (agents in specific pods), or agent (specific agents only)
Parameters	Named placeholders with defaults and descriptions, filled at runtime
Category	Organizational grouping: workflow, coding, review, safety, or custom
Priority	Higher priority skills are injected earlier in the system prompt
System skills	Built-in skills that ship with Nova; visible but not editable

How skills are resolved

When an agent runs, the skill resolver collects all applicable skills:

1. Global skills — always included
2. Pod-scoped skills — included if the agent belongs to a matching pod
3. Agent-scoped skills — included if directly assigned to the agent

Skills are ordered by priority (highest first) and formatted as an ## Active Skills section in the agent’s system prompt. The result is cached with a 30-second TTL since skills change rarely.

Example skill

Name:     code-review-checklist
Category: review
Scope:    pod (assigned to "Code Generation" pod)
Content:
  When reviewing code, check for:
  1. Error handling — are all error paths covered?
  2. Security — any injection vectors, credential exposure?
  3. Performance — unnecessary loops, missing indexes?
  4. Testing — are there tests for the changes?
  5. {{custom_check}}

Parameters:
  - custom_check (default: "Documentation — are public APIs documented?")

Planned API

Method	Path	Description
GET	/api/v1/skills	List all skills
POST	/api/v1/skills	Create a skill
PATCH	/api/v1/skills/{id}	Update a skill
DELETE	/api/v1/skills/{id}	Delete a skill
PUT	/api/v1/skills/{id}/pods	Set pod assignments
PUT	/api/v1/skills/{id}/agents	Set agent assignments

Rules — Declarative behavior constraints

Rules define behavioral boundaries for agents. They complement the Guardrail Agent’s built-in checks (prompt injection, PII, credential leak, spec drift) with user-defined policies that can be enforced both before and after tool execution.

Two enforcement paths

Enforcement	How it works	When
Soft	Rule text injected into the Guardrail Agent’s system prompt; compliance checked as part of normal LLM review	Post-execution
Hard	Regex pattern matched against tool calls before execution; blocks the call if matched	Pre-execution

A rule can use both enforcement — the pattern blocks immediate violations, and the Guardrail Agent catches subtle ones.

Actions

When a hard rule matches, the configured action determines what happens:

Action	Behavior
block	Return an error to the LLM; the tool call is not executed
warn	Execute the tool but log a warning
require_approval	Pause execution and wait for human approval

Planned seed rules

Nova will ship with three built-in system rules:

Rule	Enforcement	Action	Description
no-rm-rf	Hard	Block	Prevent recursive force delete commands
workspace-boundary	Soft	Block	Keep agents within the designated workspace
no-secret-in-output	Soft	Block	Prevent API keys and secrets in agent responses

System rules are visible and can be disabled, but cannot be deleted.

Rule properties

Property	Description
Name	Unique identifier
Rule text	Human-readable description of the constraint
Pattern	Regex for hard enforcement (optional)
Target tools	Which tools the pattern applies to (null = all)
Enforcement	soft, hard, or both
Action	block, warn, or require_approval
Scope	global, pod, or agent
Category	safety, quality, compliance, or workflow
Severity	low, medium, high, or critical

Planned API

Method	Path	Description
GET	/api/v1/rules	List all rules
POST	/api/v1/rules	Create a rule
PATCH	/api/v1/rules/{id}	Update a rule
DELETE	/api/v1/rules/{id}	Delete a rule (system rules cannot be deleted)

Dashboard pages

Skills page

• List all skills with scope badges (global / pod / agent)
• Create and edit with a content editor (markdown-capable textarea)
• Parameter definition UI (name, default, description)
• Pod/agent assignment via multi-select
• System skills shown as non-editable
• Enable/disable toggle

Rules page

• List all rules with enforcement type and severity badges
• Create and edit with a regex pattern tester (live validation)
• Tool targeting (select which tools the rule applies to)
• System rules shown as non-deletable
• Enable/disable toggle
• “Test rule” button — paste a sample tool call and see if the rule matches