Skip to main content
Every skill is defined by a SKILL.md file with YAML frontmatter and a system prompt body. The frontmatter controls execution behavior. The body defines quality standards, output format, and task-specific rules.

​
Frontmatter reference

---
name: cold-email
description: Write personalized B2B cold outreach emails
model: kimi-k2
maxTurns: 8
tools:
  - web_search
  - scrape_website
  - lookup_leads
  - search_memories
  - save_content
tags:
  - sales
  - outreach
context: fork
---
FieldTypeDescription
namestringSkill identifier used for routing
descriptionstringPlain-language description for intent detection
modelstringkimi-k2 for complex reasoning, fast for lightweight tasks
maxTurnsnumberMax agent loop iterations (6–12 typical)
toolsarrayExplicit list of tools this skill can access
tagsarrayUsed for memory tagging and activity logging
contextstringinline (main thread) or fork (isolated child agent)

​
Execution modes

Inline mode (context: inline) The skill prompt is injected as a user message into the main conversation. The skill runs in the same thread. Output appears inline. Used for short, iterative tasks where you might want to follow up. Fork mode (context: fork) The skill runs as an isolated child agent with its own token budget. The main conversation thread doesn’t accumulate the skill’s internal tool calls. Once the skill finishes, only the final result is returned to the main thread. All intermediate messages are discarded. Fork mode is used for long research tasks, heavy tool use, or anything where you don’t need to see the internal work — just the output.

​
Tool access

Each skill has an explicit tools list. Tools not in the list are not accessible during that skill run. This prevents skills from taking actions outside their defined scope. Common tool assignments by category:
CategoryCommon tools
Researchweb_search, web_search_multiple, scrape_website, browser_action, search_memories
Lead gensearch_people, search_companies, enrich_lead, verify_email, save_lead, create_task
Salesweb_search, lookup_leads, search_memories, save_content, gmail_send_email
Contentweb_search, web_search_multiple, search_reddit, save_content
Opsget_pipeline_stats, get_tasks, lookup_leads, search_memories

​
The instruction body

Below the frontmatter, the skill’s .md body contains the system prompt — task-specific instructions, quality gates, output format requirements, and banned patterns. For example, the cold-email skill body defines:
  • The 6-point quality gate (checked before returning)
  • The required email structure (specific reference → pitch → one CTA)
  • Banned words and phrases
  • The self-scoring loop (draft → score → revise if below threshold → return if passed)