Skip to main content
When you ask NIOM to do something, it doesn’t throw every tool at the problem. It routes your request through a Skill Tree — a graph of tools organized by domain — and selects the 4-8 tools most likely to succeed. This is what makes NIOM faster, cheaper, and more accurate than agents that load all tools for every request.

​
How it works

Your message
  → Skill Tree analyzes intent
    → Routes to a domain (Code, Research, Creative, etc.)
      → Selects focused tool subset (4-8 tools)
        → Agent executes with just those tools
1

Classify domain

A fast model (~200ms) reads your message and maps it to one of NIOM’s built-in skill domains:
DomainExample requests
Code”Refactor this module”, “Find bugs in my project”
Research”Research quantum computing”, “Summarize this paper”
Creative”Write a blog post about AI”, “Draft an email”
Business”Analyze this spreadsheet”, “Create a project plan”
Personal”Set a reminder”, “Organize my files”
General”What’s the weather?”, “Tell me a joke”
2

Build a Skill Path

Based on the domain, NIOM constructs a Skill Path — a ranked list of tools weighted by:
  • Domain relevance — which tools belong to this skill domain
  • Co-occurrence data — which tools tend to succeed together (learned from your usage)
  • Query embedding similarity — how close your request is to each tool’s capabilities
The result is a focused subset of 4-8 tools instead of the full 15+.
3

Detect execution mode

Before executing, a lightweight scanner checks for scheduling intent:
  • “do this every morning” → creates a recurring background task
  • “research deeply” → routes to deep research mode
  • “monitor my folder” → creates a one-shot observation task
This happens automatically — you don’t need to manually choose “background task” vs “direct response.”
4

Execute with the right tools

The agent runs with only the selected tools. Fewer tools means:
  • Faster — less context for the LLM to process
  • Cheaper — smaller prompts = lower token cost
  • More accurate — the model isn’t confused by irrelevant tools

​
Built-in skill domains

NIOM ships with 6 domain packs. Each pack defines which tools are available:
File operations, shell commands, code search, editing. Activated when your request involves coding, debugging, or file management.Tools: readFile, writeFile, editFile, searchFiles, listDirectory, runCommand
Web search, deep research, URL fetching. Activated for information gathering, analysis, and synthesis.Tools: webSearch, deepResearch, fetchUrl, readFile, writeFile
Content creation tools. Writing, drafting, editing. Activated for blog posts, emails, documents, and creative work.Tools: writeFile, readFile, webSearch, fetchUrl, deepResearch
Data analysis, planning, and organization. Activated for spreadsheets, project plans, and business tasks.Tools: readFile, writeFile, webSearch, listDirectory, runCommand
Notifications, reminders, file organization. Activated for personal productivity tasks.Tools: notifyUser, listDirectory, writeFile, readFile, runCommand
The catch-all domain. System info, basic questions, quick tasks that don’t fit a specific domain.Tools: systemInfo, webSearch, readFile, runCommand

​
The co-occurrence graph

As you use NIOM, it tracks which tools succeed together. If webSearch → fetchUrl → writeFile frequently co-occur in your research tasks, the Skill Tree learns this pattern and pre-selects all three when you ask for research. This data is stored locally at ~/.niom/skill-graph.json. It’s not sent anywhere — your usage patterns stay on your machine. Over time, tool selection becomes personalized to your workflow.

​
Schedule detection

NIOM’s Skill Tree includes a 2-tier schedule detection system: Tier 1 — Pattern matching (zero latency):
  • Detects obvious patterns: “every day”, “weekly”, “in 2 hours”
  • Triggers immediately — no LLM call needed
Tier 2 — LLM validation (conditional):
  • Fires only when the query is ambiguous or combines temporal language with actions
  • Uses the fast model to confirm whether scheduling is intended
  • Only applies if confidence > 0.7 — prevents false positives
If scheduling intent is detected, the request is automatically routed to the background task system instead of trying to execute everything in one shot.

​
Seeing it in action

You don’t need to think about the Skill Tree — it works silently. But if you’re curious, you can see it:
  1. Console logs — the sidecar logs which domain was selected, which tools were loaded, and why
  2. Skill graph file — check ~/.niom/skill-graph.json to see co-occurrence data
  3. Task creation — when the Skill Tree detects scheduling intent, it creates a task you can see in the Tasks panel
The Skill Tree is designed to be invisible when it works well. The best indicator is that NIOM consistently picks the right tools without being told which ones to use.