verify-claims

Map the terrain before designing: locate, trace, pattern, constrain. Used before any substantive design or planning step. Output is mostly for the agent's grounding; only surprises and hard constraints surface to the user.

Get 1-3 independent second opinions from fresh Claude reviewers to strengthen your reasoning before presenting to the user.

Black-box verification of any agent output: code, documents, presentations, reports, configs. Examine from the consumer's perspective and show evidence.

Surface what I think we're doing before acting on it. Produces a structured manifest exposing every silent commitment so the user can correct course before any artifact is produced.

Honest self-assessment skill that surfaces incomplete work, missing verifications, bugs, and improvement ideas before shipping.

Write a durable handoff for work that's about to ship: structured PR descriptions, summaries, memos, or memory notes. Tells the next reader what was built, why, what changed, and what's next, without forcing them to dig through the diff.

Route Claude Code's auto-created git worktrees through worktrunk so background and --worktree sessions inherit project hooks (env symlinks, installs, port allocation).

Before flipping a previously-stated position under pressure, restate its strongest form and the goal it served, then check whether the trigger refutes the core on the same goal, not just a decorative argument or a different axis.

Write the implementation spec (context, approach, files to modify, utilities to reuse, and verification) that drives execution after design is chosen.

Systematically identify how a plan or design could fail before committing to it.