菜单
Check finished work with a fresh-context subagent instead of self-critique when using Claude Fable 5. Use when designing multi-agent harnesses, verifying long-run output, or any time self-review keeps missing its own mistakes -- the maker should not grade its own work.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use before any approved Mythos-style defensive cybersecurity run, or when reviewing a Fable/Mythos security workflow. Requires authorization, bounded targets, private evidence handling, independent validation, and a patch/report owner before the model explores vulnerabilities.
Stop Claude Fable 5 from over-planning, re-deriving settled facts, or listing options it won't pursue. Use in interactive sessions where turns feel slow on simple asks, or agentic sessions at high/xhigh effort where the model gathers more context than the task needs.
Keep Claude Fable 5 running to completion in unattended pipelines -- no stalling on a stated intent without the matching tool call, no mid-run permission questions, no self-imposed session splits from context-budget worry. Use for overnight runs, scheduled jobs, and CI agents where nobody is watching to type "continue".
Pick and re-evaluate the output_config.effort level for a Claude Fable 5 workload. Use when designing a pipeline that mixes routine and hard tasks, when a call feels slow or expensive for what it's doing, or when porting an Opus 4.7/4.8 harness that defaulted to xhigh.
Bind every progress claim Claude Fable 5 makes during a long or autonomous run to an actual tool result from that session. Use for multi-hour agents, overnight pipelines, CI runs, or any workflow where a status update has previously claimed work that wasn't actually done.
Maintain a file-based lesson memory that Claude Fable 5 reads and updates across sessions. Use for recurring agents (daily jobs, long projects, team assistants) where the same correction keeps being re-made, or when bootstrapping a new agent from past session history.
| name | verifier-subagent |
| description | Check finished work with a fresh-context subagent instead of self-critique when using Claude Fable 5. Use when designing multi-agent harnesses, verifying long-run output, or any time self-review keeps missing its own mistakes -- the maker should not grade its own work. |
Fable 5 dispatches and sustains parallel subagents dependably enough to make this practical at scale, and the pattern is worth defaulting to: a separate subagent with a clean context outperforms self-critique at catching defects, because it doesn't share the maker's assumptions or blind spots.
The agent that produced the work does not grade it. Give the verifier the specification and the output -- never the maker's reasoning or narration about what it did. A verifier that reads "I fixed the race condition by adding a lock" will look for a lock; a verifier that only reads the diff will look for a race condition.
Split out a subtask when it is independent of your current context, large enough to amortize the handoff, and specifiable in a few sentences plus file pointers. Don't delegate tightly coupled edits -- coordination cost exceeds the parallelism gain.
For a long-running build, run verification at a defined interval (every N components, every X hours), not once at the end. A defect caught at component 3 is cheap; the same defect caught after component 40 depend on it is not.
A subagent brief needs exactly: the goal (one sentence), inputs (paths or data), a checkable definition of done, constraints (what not to touch), and where to write results. If you can't fill in all five, the task isn't ready to delegate.
