← ClaudeAtlas

spec-intakelisted

Turn a raw idea into a validated, dependency-ordered spec and file it as a batch of GitHub issues that feed the batch lane. You run the conversation and the decomposition; deterministic `orchestrator spec` code validates, plans, and files. The front door above intake — issues out, not code.
cperler/sous · ★ 1 · AI & Automation · score 67
Install: claude install-skill cperler/sous
# Spec intake — idea → validated spec → dependency-ordered issues You are the **front door**. A run starts from an already-written issue; nothing upstream turns *an idea* into well-scoped, dependency-ordered issues. That's this skill. You own the conversation and the decomposition; the deterministic `orchestrator spec` commands own validation, ordering, and filing. You never file issues by hand — `spec file` does it, in dependency order, translating local ids to real issue refs. **Filing is outward-facing: the human confirms the plan before anything is created.** ## Constants - `PROJECT` = the project-config module/dir supplying the task source (e.g. `adapters.project.selfhost`). - Command shape: `uv run orchestrator [--project "$PROJECT"] spec <validate|plan|file> <file>`. ## The flow 1. **Interrogate the idea.** Pin down the goal, the hard constraints, and the explicit non-goals. Ask the 2–3 questions that most change the decomposition; don't interview. 2. **Decompose.** Break it into **small, independently-shippable tasks**, each with its own acceptance criteria and explicit dependencies. A good task is one PR's worth of work with a clear "done" test. Make dependencies real (t2 needs t1's output), not incidental ordering. Prefer more, smaller tasks over few big ones. 3. **Write the spec JSON** to a file (schema: `orchestrator/schemas/spec.json`): ```json { "title": "...", "summary": "goal / constraints / non-goals", "tasks": [ { "id":