preflightlisted
Install: claude install-skill muzalee/claude-atelier
A plan is a set of claims about a repository: this file exists, that function takes these arguments, this dependency is installed, this command runs. Every one of those claims is checkable, and a plan is only as good as the claims it got right.
This skill checks them, and reports what would break — before a build hits the wrong assumption at step 4 and quietly improvises the rest. That improvisation is the failure this exists to prevent: the plan was reviewed and agreed, the code that shipped was something else, and nobody saw the divergence because it happened mid-execution.
**This skill fixes the plan; it never touches the code.** It corrects what has one right answer, asks about what has several, and stops only at decisions that are genuinely the user's — so what comes out the other side is a plan `/build` can run, not a list of chores handed back.
**Why it edits the plan instead of writing a report.** This is the step where a human is present. `/build` can be told `go all` and left alone, and `/ship` runs the whole pipeline with nobody watching — but neither can answer a question, and `/ship` rule 2 halts the entire run on one. Every question preflight resolves *now* is a halt that does not happen later. That only works if the answers land **in the plan file itself**: a report beside the plan is a second document the unattended build never reads, and the build would improvise past exactly the step the report was warning about.
So the output of a good pass is not a doc