Edit a .spec.ts file to update the project's instruction files. The spec is the source of truth — CLAUDE.md and AGENTS.md are compiled build artifacts that must not be edited directly.
Arguments
$ARGUMENTS — What the user wants to change. Examples:
- "add a rule about always using the custom logger"
- "update the architecture section"
- "point the auth rule at src/services/auth.ts"
- "add npm run lint to commands"
- "change the testing guidance"
Instructions
Step 1: Find the Spec
Look for spec files in the repo root:
CLAUDE.md.spec.ts— source for CLAUDE.mdAGENTS.md.spec.ts— source for AGENTS.md- Any
*.spec.tsmatching instruction files
If no spec exists: if there's a hand-written CLAUDE.md, suggest the
adopt-spec skill; otherwise suggest npx vigiles init to scaffold one.
Step 2: Read and Understand the Spec
Read the spec file. It's a TypeScript file that exports a claude() call with these fields:
import { claude, enforce, guidance, check, every } from "vigiles/spec";
export default claude({
// Optional: output target (defaults to "CLAUDE.md")
target: "CLAUDE.md",
// or multi-target:
// target: ["CLAUDE.md", "AGENTS.md"],
// Prose sections — become ## headings in compiled output
sections: {
positioning: "What this project does...",
architecture: "How the codebase is structured...",
},
// (No `keyFiles` — deprecated; see "A pointer to a file" in Step 3.)
// Commands verified against package.json
commands: {
"npm run build": "Compile the project",
"npm test": "Run all tests",
},
// Rules — three types
rules: {
// enforce() — backed by a linter rule, verified to exist AND be enabled
"no-any": enforce(
"@typescript-eslint/no-explicit-any",
"Use unknown and narrow with type guards.",
),
// check() — filesystem assertion run by vigiles
"test-pairing": check(
every("src/**/*.service.ts").has("{name}.test.ts"),
"Every service must have tests.",
),
// guidance() — prose only, no enforcement
"research-first": guidance("Google unfamiliar APIs before implementing."),
},
});
Step 3: Make the Changes
Based on what the user asked for:
Adding a rule (this absorbs the old generate-rule skill):
- Classify the rule from the request:
enforce()— a linter rule can back it. Check the project's linter configs (ESLint, Ruff, Clippy, Pylint, RuboCop, Stylelint) for a matching rule; also consider an architectural tool (ast-grep, Dependency Cruiser, Steiger). If uncertain whether a rule exists, ask the user rather than guessing.check()— a filesystem structural convention ("every X needs a Y"). Only for file-pairing; never for code content.guidance()— can't be mechanically enforced (subjective conventions, process rules, migration context).
- For
enforce(): use the real linter rule name (e.g.eslint/no-console,@typescript-eslint/no-explicit-any,ruff/T201). - Add to the
rulesobject with a kebab-case key derived from the intent, preserving alphabetical order if the existing rules are alphabetical. Import any new builders needed (e.g.checkandeveryfor the firstcheck()).
Updating a section:
- Edit the string in
sections. Sections are plain strings or tagged template literals withfile(),cmd(),ref()for verified references - Do NOT add
#or##headers inside sections — they break the document structure
Adding a command:
- Add to
commands. The compiler verifies it againstpackage.jsonscripts. - Keep the description to one line (aim for 120 characters, never exceed ~200): the instruction file is loaded on EVERY request.
A pointer to a file (what used to be a "key file"):
- 🔴 Never add a
keyFilesentry — not one, not even to a spec that already has the field. It is deprecated (vigiles lintreports it asdeprecated-key-files) and removed in the next major: it compiles a file map into the instruction file loaded on every request, and the agent finds files itself. - Asked to add a file? Place it by the first of these that fits, and say why:
- the file's own header, the docs or the path already say it → add nothing;
- a fact about that one file → write it in the file's header comment;
- a rule an agent must follow while working in one folder → that folder's
nested
CLAUDE.md(its own<folder>/CLAUDE.md.spec.ts, compiled); - only if nothing else would lead an agent there → a one-line pointer,
${file("src/services/auth.ts")}inside aprosesection where the decision that needs it is written — compile verifies it exists, exactly as it verified akeyFilespath. (guidance()takes a plain string, so a path typed there is not verified.)
- Opened a spec that still has
keyFiles? Propose removing it, in the same reply: go through the entries with the same four steps (the full recipe isdocs/rules/deprecated-key-files.mdin the vigiles repository; a path that no longer exists is deleted) and show, per entry, where it goes and why. Where each entry goes is the user's call; apply it once they confirm.
Step 4: Compile
After editing the spec, run:
npx vigiles compile
This regenerates the compiled instruction file(s). Review the output for any errors:
stale-file— afile()(or deprecatedkeyFiles) path doesn't existstale-command— a command isn't in package.jsoninvalid-rule— a linter rule doesn't exist or is disabledsection-has-header— a section contains#headers (break into separate named sections)
Step 5: Verify
npx vigiles lint
If the vigiles plugin is installed (/plugin marketplace add zernie/vigiles then /plugin install vigiles@vigiles, or npx vigiles init), the PostToolUse hook recompiles automatically after you save the spec.
Important
- Do what's asked; don't silently escalate enforcement. Add the rule the user asked for. If making it an
enforce()would need a linter-config edit or a plugin install (a cost), or if it could fail a clean CI, say so and let the user choose — don't change linter config or flip on strict /workflowgating on your own. A pure win (a rule that's already enabled) you can just apply. Thestrengthenskill ownsguidance()→enforce()upgrades. - Never edit CLAUDE.md or AGENTS.md directly — they have a vigiles hash comment and are build artifacts
- The spec is TypeScript — you get type checking, autocomplete, and verified references
enforce()rules are verified — the compiler checks the rule exists AND is enabled in your linter config- Sections must not contain
#or##headers — use separate named sections instead
