Docs are for humans.
Skills are for agents.
Structured instructions. Zero hallucinations.
01
Step 01
Agent fetches skill
curl llms.txt | or npx skills add Agent identifies what it needs to build and pulls the right skill file
02
Step 02
Reads instructions
{ pitfalls, steps, verify } Gets structured steps, working code, known pitfalls, and verification checks
03
Step 03
Builds correctly
✓ deployed & verified Executes with zero hallucinations because every detail is precise and tested
What changes
Before
◌ Agent guesses canister IDs
◌ Hallucinated Motoko syntax
◌ Outdated CLI commands
◌ No idea about cycles or fees
◌ Broken deploys, wasted time
◌ Human has to fix everything
→
After
◆ Exact canister IDs from skill
◆ Tested, working code templates
◆ Current icp-cli commands, verified
◆ Fees and cycles handled correctly
◆ Deploys work first try
◆ Agent ships autonomously
Anatomy of a skill
Any agent, any framework
Claude
Skills as context
ChatGPT
Function calling
Cursor
Rules files
Devin
Knowledge base
Copilot
Custom instructions
Windsurf
Cascade context
Claude Code
SKILL.md files
OpenCode
Remote instructions
OpenClaw
Skills marketplace
skills.sh
npx skills add
The first chain agents can actually build on.
Structured skill files that prevent hallucinations. Fetch via llms.txt, raw GitHub URLs, or the skills.sh CLI.
npx skills add dfinity/icskills