Interview yourself into an agent spec

Most agents fail because they were never specified, just vibed. 'I want an agent that finds deals' is a wish, not a build. The move that fixes it costs ten minutes: have Claude Code interview YOU, one question at a time, and turn the wish into a spec. Six things every agent needs pinned down before a line of code: the INPUT (what it reads), the TRIGGER (what makes it run, a schedule, a new row, a manual kick), the OUTPUT (what it produces and where it lands), the GUARDRAILS (what it must never do), the COST CAP (a hard ceiling on runs, tokens, or spend, so a loop can't drain your account), and the DEFINITION OF DONE (how you'll know a run was good). Make Claude ask, don't let it assume, and you'll catch the ambiguity now instead of at 2am when the agent has emailed the wrong list. The spec is the artifact; the code is downstream.