claude-dotfiles

Call the Skill tool twice, for "grilling" and "domain-modeling".

One adversarial pass before the spec is published

When the interview has settled and before the thinking goes to /to-spec, spend one sub-agent on
attacking it. Give a fresh context the plan as it stands, plus the CONTEXT.md entries and ADRs it
rests on, and tell it to argue that the plan fails: the assumption nobody checked, the case the
design has no answer for, the ADR whose rejected alternative was dismissed too fast. Its tools are
Read, Grep, Glob and Bash — no Edit, no Write — so it argues instead of quietly rewriting.

The context that ran the interview cannot do this: it produced the plan and will defend it.

Put each objection back to the user as a grilling question and settle it. Changing the plan is free
here and expensive once the spec is published.