Skill

Reply in the contract language.

Put skill full in architecture.hexum. On the next MCP call the agent asks hexum_skill, writes SKILL.md if it is missing, and answers in cause / fix / do / check. Code, errors, and names stay exact.

Normal verbose

The reason this fails the architecture
check is that App.tsx lives in the UI
layer and imports orders from data.
I'd recommend calling the services
layer instead so the UI does not talk
to data access directly.

Hexum full

cause src/ui/App.tsx
  import ../data/orders

fix
  call services

check
  rule no_ui_to_data

Modes

Mode	Reply	Read
`lite`	Full sentences. No filler.	Shrink wide context.
`full`	Hexum blocks. Default.	Shrink, expand the symbol you edit.
`ultra`	One cause, one fix, one check.	Expand only the symbol you splice.

Set in architecture.hexum, or call hexum_skill with intensity. off restores prose.

What this does not claim

The skill shrinks what the agent says. The measured cuts on MBPP (~85%) and HumanEval (~89%) are from hexum_shrink on input, same tests passing. No output-token study is published. Commits, PRs, comments, and docs stay normal English.

Install

Connect MCP, then set the mode in the architecture file. The agent writes the skill file. You do not copy it by hand.

# architecture.hexum
skill full

layer ui
  match src/ui/**

Ask the agent to turn Hexum on, or it calls hexum_skill itself when it sees the skill line. MCP returns the body; the agent writes .claude/skills/hexum/SKILL.md (Cursor: .cursor/skills/hexum/, Codex: .codex/skills/hexum/, Grok: .grok/skills/hexum/) if that path is missing.

To switch: hexum_skill with intensity lite, full, ultra, or off. That upserts the line in architecture.hexum too.

Mint a hexum_live_ key and paste the snippet on Agents.

Get API key Docs