Blind Spot Pass

Your prompt is the map; the real project is the territory. The gap between them is full of decisions you never made — and an agent will not stop and ask. It picks something plausible, builds on it, and hands you a finished thing built on a guess. Blind Spot Pass runs the step people skip: before you ask the agent to build the thing, it finds the parts of the thing you forgot to mention. Give it a rough plan, spec, ticket, or prompt — rough is fine, the unfinished version is the one worth examining — and it reads it the way an agent would execute it, literally, without the context that lives only in your head.

It returns the silent assumptions, each stated as the decision you didn't make plus the default an agent would most likely pick (you didn't say whether deleted items are recoverable; an agent will implement a hard delete); a ranking by what a wrong guess actually costs, separating rebuild-class decisions like the data model, the auth boundary, and who-can-see-what from the tweaks that cost minutes; a short ordered list of decisions to make now, phrased as concrete choices with a suggested default, because a choice takes ten seconds and an open question takes a meeting; the gaps that are explicitly safe to leave open, so the list stays short enough to actually answer; and the one question that most shapes what gets built.

The download includes three reference files: the blind-spot taxonomy, a decision-elicitation worksheet, and a worked sample that takes a fifty-one-word dashboard plan and finds seven unspecified decisions. It examines the plan you provide — it doesn't inspect your codebase or run anything, and it surfaces decisions rather than making them. Works with Claude Code, Cursor, Codex CLI, Gemini CLI, and any SKILL.md agent.