Use when you have a free-form draft prompt and want it rewritten into Claude Opus 4.6 or 4.8 conventions before pasting it into a new planning or execution session. Trigger phrases — "align prompt", "rewrite for opus", "tune prompt for claude". NOT for planning itself, code implementation, or validation. Pre-planning / pre-execution only.
Post-session validation loop — three-wave checks (Deterministic / Rule-based / Probabilistic) before commit. Use before every git commit. Writes .validation_passed sentinel with waves_passed field on success. NOT for: ad-hoc test runs, code review, or implementation work — only the Pipeline Gate between implement and commit.
Invoke the plan-reviewer agent using the canonical reference prompt from docs/plan-reviewer-design.md. Use when reviewing a plan in a fresh Claude Code session (after the plan was created in a prior session) and you want adversarial review (checklist + Elegance Gate + handoff plan) without manually copy-pasting the reference prompt. Accepts the plan path as an argument. NOT for: same-session plan reviews (just invoke the agent directly with the plan in context), reviewing implementations that already shipped (use /multi-review or /session-critique), or spec audits (the agent has a separate Spec Audit mode invoked differently).
Author a CONTEXT.md routing file for a specified project subdirectory using the canonical ICM anatomy (.claude/rules/context-md-anatomy.md). Use when adding a new high-traffic area whose routing logic the root CLAUDE.md cannot describe economically. Skip for trivial directories. NOT for: editing existing CONTEXT.md files, authoring CLAUDE.md, or documenting project architecture — only new L1 routing files.
Autonomously improve a generated paper via REVIEWER_MODEL (default o3) xhigh review → implement fixes → recompile, for 2 rounds. Use when user says "改论文", "improve paper", "论文润色循环", "auto improve", or wants to iteratively polish a generated paper.
Draft LaTeX paper section by section from an outline. Use when user says "写论文", "write paper", "draft LaTeX", "开始写", or wants to generate LaTeX content from a paper plan.
Hypothesis-first interpretation of eval results. Use after an eval batch completes, after /post-eval populates analyses, or when preparing paper narrative from result JSONs. Phase 1 requires a prior hypothesis (prevents post-hoc rationalization); Phase 2 compares to actual data.
Zero-context verification that every number, comparison, and scope claim in the paper matches raw result files. Uses a fresh cross-model reviewer with NO prior context to prevent confirmation bias. Use when user says "审查论文数据", "check paper claims", "verify numbers", "论文数字核对", or before submission to ensure paper-to-evidence fidelity.