Skip to main content
A skill is a focused AI sub-agent with its own model, tool set, token budget, and instructions. You don’t configure skills — you describe what you need and Ultron routes to the right one.

​
How skill execution works

You: "Write a cold email to the Head of Growth at Linear."
    ↓
Chat detects intent → routes to cold-email skill
    ↓
Skill loads: model=kimi-k2, maxTurns=8,
  tools=[web_search, scrape_website, lookup_leads, save_content]
    ↓
Context built:
  ├── Your business profile (ICP, voice, competitors)
  ├── Past cold emails in memory (what worked, what didn't)
  └── Relevant lessons (quality gate failures, style corrections)
    ↓
Kimi K2 runs the skill agent loop:
  ├── web_search("Linear Head of Growth 2026")  ← parallel
  ├── scrape_website("linear.app")              ← parallel
  └── lookup_leads("Linear")
    ↓
Email drafted → 6-point quality gate applied → passed
    ↓
Result returned to chat
Memory saved: ["skill-run", "sales", "cold-email"]
Activity logged

​
Concurrency

Max 2 skills run simultaneously per session. If you invoke a third while two are active, it queues and starts the moment a slot opens. Rate limits (429 responses) trigger exponential backoff (1s → 2s → 4s) with up to 3 retries.

​
Execution modes

ModeHow it worksBest for
InlineSkill continues in the main conversation threadShort tasks, iterative work
ForkSkill runs as isolated child agent, results returnedLong research, heavy tool use

​
Skills by category

Research

Competitive analysis, company deep-dive, pricing research, meeting prep.

Lead Generation

Find decision makers, ICP matching, lead scoring, trigger events.

Sales

Cold email, follow-up sequences, objection handling, pre/post call.

Content

LinkedIn posts, social content for all platforms.

Operations

Morning briefing, pipeline review.

How skills work

Invocation, progress streaming, memory auto-save.