メニュー
Use as a brief local routing step before non-trivial planning, execution, review, refactoring, debugging, or design when a shallow reading could change scope, risk, or user-visible behavior. Keep it silent when intent is clear; it does not imply an artifact or subagent.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Use for an actual ready, non-draft GitLab merge request review, re-review, or approval decision when current discussions, diff, CI, scope, findings, and approval state must be checked. Do not trigger for casual MR discussion, draft/WIP work, GitHub PRs, or local code review.
Use at conversation start, after context recovery, and for non-trivial engineering work to select exactly one primary workflow path plus one task-wide assurance and verification budget. Loading this router does not create artifacts, launch agents, or activate every workflow module.
Use only when creating, resuming, inspecting, updating, completing, or writing artifacts for an authorized .workflow plan or audit run through the bundled Workflow MCP tools. It is an operations contract, not a reason to create workflow artifacts.
Use when an authorized, sufficiently clear change needs a durable decision-complete plan under .workflow because execution is multi-step, risky, resumable, or explicitly requested. Do not create a plan for fast local work or read-only discussion.
Use at conversation start, after compaction or context recovery, and before Agent Memory reads or writes to decide whether durable global or project knowledge can materially change the task. Keep startup reads focused, and do not mutate memory during read-only or no-edits work unless the user explicitly asks to remember something.
Use after intent routing only when unresolved requirements, product intent, architecture, constraints, or tradeoffs can materially change scope, risk, acceptance criteria, or user-visible behavior. Inspect discoverable facts first and avoid exhaustive questioning.
| name | intent-gate |
| description | Use as a brief local routing step before non-trivial planning, execution, review, refactoring, debugging, or design when a shallow reading could change scope, risk, or user-visible behavior. Keep it silent when intent is clear; it does not imply an artifact or subagent. |
Determine what the user actually wants, whether the obvious reading is safe, and which primary path and assurance profile fit. This is a brief local routing decision, not optional ceremony and not an automatic subagent step.
simple, medium, complex, very_complex) for execution shape and assurance (fast, standard, assurance, explicit audit) for risk.Use L0 for content/config-only changes, L1 for a localized behavior change, L2 for a cross-module contract or reusable abstraction, and L3 for a new subsystem/platform layer. Do not silently raise the class. A new API/schema/protocol/shared abstraction/layer outside the expected envelope requires a narrower alternative, explicit justification, or a material user question.
Before presenting any material plan, architecture decision, or scope-expanding solution, take one local second look from existing context: compare it with the request and envelope, remove unsupported scope, and choose KEEP, SHRINK, ASK, or REPLAN. Do this silently for chat-only work; do not add an agent or discovery pass.
When confidence is high, continue without printing a gate report. When a decision must be exposed, report only intent, assumption/question, assurance, and next path.
Use read-only workflow_intent_reviewer only when all are true:
Routine implementation, bounded reading, clear corrections, and multiple applicable skills never justify this reviewer. If unavailable, continue locally unless the user explicitly requires independent intent review.
Stop the gate as soon as the workflow path is safe to choose.