ghostwriter AGENTS.md

Repository instructions

Three independently installable Agent Skills. Keep the repository local-first, dependency-free at runtime, and safe for personal writing data.

Before editing

• Read the nearest SKILL.md and every reference needed for the change.
• Keep each skill self-contained. A skill must not reference files outside its own directory at runtime.
• Use fictional fixtures only. Never copy a personal profile, export, held-out response, credential, or eval run into Git.
• Preserve the boundary between cases.jsonl and references.jsonl.

Implementation

• Scripts are TypeScript run directly by Node's type stripping, so keep every file to erasable syntax (no enums, namespaces, or parameter properties).
• Use Node.js standard-library modules. Do not add a runtime dependency for work the standard library handles clearly.
• Spawn local agents with argument arrays and stdin, never through an interpolated shell command.
• Keep deterministic data work in scripts and language judgment in the invoking agent.
• Use plan, validate, execute for profile replacement and other multi-file writes.
• Keep tone rules in private profiles. Runtime instructions must not hard-code one person's stylistic habits.
• Persistent user data belongs under GHOSTWRITER_HOME, never inside an installed skill.

Verification

Before declaring a change complete:

npm test
uvx --from skills-ref agentskills validate skills/ghostwriter
uvx --from skills-ref agentskills validate skills/train-ghostwriter
uvx --from skills-ref agentskills validate skills/evaluate-ghostwriter

Model-backed smoke tests are opt-in. CI and ordinary unit tests must use fictional fixtures and stub executables.