Skip to content

Scenario — the language switch

A Spanish-speaking operator wants ways in Spanish. This is the path the adopter-run design exists to serve: the operator asks, the ways-localize skill confirms the cost, and only on consent does ways localize itself.

How it plays out

sequenceDiagram
    autonumber
    participant Op as Operator (Spanish)
    participant CC as Claude Code
    participant L as ways-localize skill
    participant W as agent-ways · language = auto (English)
    rect rgba(217,119,6,0.12)
    Op->>CC: "configura ways en español"
    CC->>L: request matches the skill's description
    L-->>Op: confirm Spanish (es), state the cost: 127 MB model, every way translated
    Op->>L: yes (consent)
    end
    rect rgba(45,125,154,0.12)
    L->>W: set language → es
    L->>L: fetch multilingual model · translate every way against the English root
    L->>L: rebuild corpus (English anchor) · ways tune locale until clean
    L->>CC: settings.json language → spanish
    end
    rect rgba(45,142,94,0.12)
    Note over Op,W: next session: ways matches in Spanish, Claude Code answers in Spanish
    end

What each move is doing

  • The operator asks; nothing asks for them. ways does not watch Claude Code's language. Setting settings.json language: spanish on its own changes Claude Code's replies and leaves ways in English mode. The request "set up ways in Spanish", in any language, is what starts localization: the ways-localize skill's description is written to match it. There is no automatic cascade from "CC is Spanish" to a model download and a translation pass.
  • The skill asks permission before the heavy part. Localization is expensive: a 127 MB model download, a translate-every-way pass and a tuning loop. The skill confirms the language against the registry, states that cost, and waits for a yes.
  • On consent, ways-localize does the work and flips the flag. It sets ways' language to es, fetches the multilingual model, translates each way's description and vocabulary against the English root, rebuilds the corpus with the English anchor, and runs ways tune locale until the Spanish layer is clean: aligned to the root, with no collisions (see 01.013.E).
  • Claude Code follows last. The skill sets Claude Code's own language to spanish and reports in Spanish. Both take effect next session.

The point

The operator gets a localized experience when they ask for one, pays its cost themselves as the beneficiary, and is the native speaker best placed to judge the result. After this, the install sits in localized steady state (01.012.E).