// the find
Klerith/fernando-skills
Spec-driven development skills for Claude Code, Cursor, Codex, and Antigravity. By Fernando Herrera.
A pair of Claude Code skills (plus ports for Cursor, Codex, and Antigravity) that formalize a spec-driven workflow: describe the problem, let Claude propose a plan, refine it into a saved markdown spec, then implement it step by step with review pauses between each step. It's for developers who use AI coding agents on nontrivial features and want a persistent decision log instead of re-explaining context every session.
The six-step procedure with human-gated status transitions (Draft → Approved) is a real safeguard — /spec-impl explicitly refuses to run against anything not marked Approved, so there's an actual checkpoint rather than a suggestion. The 'anatomy of a spec' section is concrete and opinionated about what makes acceptance criteria verifiable versus vague ('works well' is called out as unacceptable), which is more useful than most spec templates. Multi-agent support isn't just a README claim — install-to-agent.sh actually generates agent-specific artifacts (Cursor .mdc rules, Codex AGENTS.md blocks, Antigravity skill copies) rather than assuming everyone uses Claude Code.
This is a documentation-and-shell-glue project (Shell as the listed language, two SKILL.md files) — there's no test suite, so nothing verifies the installer scripts or skill logic beyond manual use. The workflow leans entirely on markdown prompts interpreted by an LLM; there's no tooling enforcing that a spec's implementation plan actually matches what /spec-impl executes, so drift between spec and code is still possible despite the stated philosophy. Status labels being 'language-agnostic free text' (any word meaning Approved works) means there's no validation catching a typo or synonym mismatch — it's convention enforced by prose, not code.