Skills

Skills are the core building blocks of PancakeSwap AI. Each skill is a Markdown file (SKILL.md) with YAML frontmatter that tells an LLM agent exactly how to perform a task.

All Skills

Skill	Plugin	Model	Description
swap-planner	pancakeswap-driver	Sonnet	Plan swaps with token discovery and deep link generation
liquidity-planner	pancakeswap-driver	Sonnet	Plan LP positions with pool assessment and APY analysis
collect-fees	pancakeswap-driver	Sonnet	Check and collect pending LP fees across V3 and Infinity positions
swap-integration	pancakeswap-driver	Opus	Integrate PancakeSwap swaps into apps using Smart Router or V2 Router
farming-planner	pancakeswap-farming	Sonnet	Plan yield farming, CAKE staking, and deep links
harvest-rewards	pancakeswap-farming	Sonnet	Check pending farming rewards and generate harvest deep links
hub-api-integration	pancakeswap-hub	Sonnet	Integrate PCS Hub quote/swap API into wallets and partner frontends

Skill Anatomy

Every skill file follows the same structure:

---
name: skill-name
description: When to activate this skill (trigger phrases)
allowed-tools: Read, Write, Edit, Bash(npm:*)
model: opus
license: MIT
metadata:
  author: pancakeswap
  version: '1.0.0'
---

# Skill Title

Content: decision guides, code templates, address tables,
validation rules, anti-patterns, and more.

Frontmatter Fields

Field	Required	Description
name	Yes	Unique skill identifier
description	Yes	Trigger phrases — when the agent should load this skill
allowed-tools	Yes	Tools the agent may use while executing this skill
model	No	Recommended LLM model (opus, sonnet)
license	No	License for the skill content
metadata.author	No	Author name
metadata.version	No	Semantic version

Design Principles

1. Self-contained — each skill has everything an agent needs (addresses, ABIs, code templates)
2. Decision-first — start with a decision guide so the agent picks the right approach
3. Correct by default — code templates use verified addresses and safe defaults
4. Anti-pattern aware — explicitly call out common mistakes with ::: danger blocks
5. Chain-aware — tables map every address and behavior to specific chain IDs
6. Testable — each skill has a corresponding eval suite under evals/suites/