System

Dispatch a foreign (non-Anthropic) model as a subagent — OpenAI Codex/GPT-5.6-Sol, DeepSeek V4-Pro, or Google Antigravity — either as a write-authorized agentic worker in a worktree or a read-only reviewer. Codex and DeepSeek explore, edit when authorized, and self-report; Antigravity is a single-packet headless reviewer. Use when you (the conductor) want to offload agentic work to another harness, get a cross-lineage second opinion, or run a correlated-blind-spot check same-lineage Claude subagents can't give. This is the dispatch/wrapping layer — how to split work across foreign models of unequal intelligence, prompt each to its level, and wrap them safely to save tokens. Also where adversarial-crosscheck's foreign lane routes; not every dispatch is a crosscheck.

Take a prompt or task the user is about to give Claude — especially a judgment they're invested in ("is this plan sound?", "review my design", "should we adopt X?", "poke holes in this", "am I fooling myself here?") — and run it through two opposing rewrites in clean contexts (a neutralised pass and a disowned-and-inverted adversarial pass), then present both for the user to reconcile. Use this whenever the user wants to pressure-test their own thinking, worries they're being flattered or agreed-with, asks to "red-team", "cross-check", "steelman against", or "sanity-check" something, or hands over work they clearly want to succeed and asks for an honest assessment. It counters sycophancy by widening coverage — it does NOT produce a single "unbiased" or "debiased" answer, and it deliberately refuses tasks it can't honestly automate.

Deep, source-graded investigation for substantial questions — casts a wide net of primary sources, grades each for quality and relevance, and works from the graded base instead of guessing. A good default for most questions worth more than a snap answer: tool/library/architecture choices, approach tradeoffs, trustworthiness or maintenance judgments, root-causing unfamiliar behaviour, or any "which way and why" that rewards real evidence.

Reviews prompt-engineering artifacts (SKILL.md files, CLAUDE.md, AGENTS.md) against a built-in checklist of effectiveness criteria for steering Claude, and proposes only substantively-justified edits with a brief rationale per change. The criteria carry a "criteria-as-of" date; when stale (>180 days), the audit still runs against the current criteria but recommends the user trigger the (separate, heavyweight) refresh procedure in a fresh max-effort session — refresh is never run inline. Explicitly designed to be token-expensive and slow; compounding value because the audited artifacts are read by Claude across many future sessions. Read after writing or substantively editing any SKILL.md, CLAUDE.md, or AGENTS.md.

The user's idiosyncratic commit-message style — `(label1 label2) terse message` with the mandatory `AI` tag for AI-authored commits, using the `.gitlabels` convention. Two-mode operation — autonomous-commit on `ai/*` branches (or with a `.claude-commit` sentinel), message-only on others. Read whenever drafting a commit message, staging a commit, or finishing a chunk of work that will produce a commit.

End-of-work hygiene pass for AI-edited code, focused on stripping the overly-verbose code comments that AIs habitually generate. The user wants ~90% of AI-added comments deleted, the remaining ~10% rewritten for brutal brevity. Read this before declaring an autonomous task complete, before drafting a commit on substantive edits, or whenever the user asks for a "review pass", "cleanup", or "wrap-up".

Machine investigation, deployment, Ansible, SSH, 1Password, Notion — read for any infrastructure/ops/server-management tasks

Writing and modifying scripts; read for any shell/bash/zx/node.js/ansible/Makefile-type tasks