← ClaudeAtlas

grill-speclisted

Finds the holes in a spec before code exists — works a fixed set of frontend dimensions for the unstated state, the missing permission case, the timezone bug, the thing that breaks at 10,000 rows. Run it on a spec draft before implementation, on an architecture decision you are about to commit to, or when a plan needs holes found rather than agreement. Findings go back into the spec.
AlpayC/claude-skill-set · ★ 0 · AI & Automation · score 60
Install: claude install-skill AlpayC/claude-skill-set
# Grill spec Every question answered here is one the implementation does not stop for. Every one missed becomes a bug report with a screenshot attached. Agreeing with a plan produces nothing its author did not already have, so be relentless: the job is to find what is missing and keep pulling until the answers stop being confident. ## How to ask **Look it up rather than asking.** Anything discoverable — how the neighbouring feature handles this, what the endpoint returns, which roles exist, what the design system offers — is yours to find, with a subagent where it is a real search. The user's time is for decisions only; asking them a question you could have answered yourself spends it on the wrong thing and teaches them the session is expensive. **Ask in batches of three to five**, and wait for answers before the next batch. Twenty questions at once get twenty shallow answers. **Order by dependency.** Some answers unlock other questions — whether errors surface inline or as a toast determines what the error copy questions even are. Ask what unblocks the most first, and hold back anything whose phrasing depends on an answer you do not have yet. **Carry a recommendation.** Every question comes with the answer you would give and why: ``` Q2 — Empty state, first use vs. filtered to nothing The mockup shows one empty state. These usually need different copy: a new user needs orientation, a filtered-out user needs a way back. → Recommend: two states. Filtered-out gets a "Fi