wiswa-mcp: Skill for Claude Code

Prose lint

Audit prose against .claude/rules/prose.md and apply the fixes.

Read that rules file in full before the first edit. This skill describes the procedure only; the rules file is the authority on what is banned.

Audit adversarially. When a sentence is defensible under a loose reading and a violation under a strict one, treat it as a violation. A borderline construction is rewritten, not excused, and doubt resolves toward the edit. Under-reporting is the failure mode to avoid. A grep hit stands until you can state which exemption in this file covers it.

Scope

Default target

Audit files with pending changes, from git status --porcelain and git diff --name-only HEAD. When the user gives paths or a diff range, use those instead.

Audit every piece of prose inside each target file:

• Markdown body text, including headings, list items, table cells, and admonitions.
• Plain comments in any language.
• Formal documentation comments (Numpydoc, JSDoc, Doxygen, Javadoc, XML doc comments). The same standard applies to them as to Markdown body text.
• User-facing string literals such as log messages, help text, error messages, and CLI output.
• Commit messages under review.

Do not edit:

• Identifiers, code, URLs, file paths, and dependency names.
• Assistant instructions. AGENTS.md, CLAUDE.md, and everything under .claude/ are out of scope, including when the user requests every file (everywhere, the whole repository, a glob that covers them). Report a hit in one of them only when the user has requested a change to that file.
• Fenced blocks and italic examples inside .claude/rules/prose.md. They exist to demonstrate violations.
• Released CHANGELOG.md sections. Only the Unreleased section is in scope.
• Quoted material from an external source. Flag it for the user instead.

Detection

Run the greps below over the target files, then review every hit by hand. Each grep over-matches by design; a noun matching a banned verb is valid, and the fix depends on the sentence. claim is valid wherever a specification defines it as a term, such as an OIDC or JWT claim. The noun project is valid and accounts for most hits on that word. Only the verb is banned, and the geometric sense of the verb is valid too. The noun ask is banned as well. Review every hit on that word rather than excusing the noun.