메뉴
Full multi-phase orchestration for one task — understand → design → implement → review, each phase backed by multi-agent workflows with adversarial verification. Use for substantive features, refactors, and audits where the user wants maximum quality over minimum tokens.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
| name | ultra |
| description | Full multi-phase orchestration for one task — understand → design → implement → review, each phase backed by multi-agent workflows with adversarial verification. Use for substantive features, refactors, and audits where the user wants maximum quality over minimum tokens. |
| argument-hint | <task> |
The user invoked /ultra: that is explicit opt-in to multi-agent workflow orchestration for this task. Call the Workflow tool as directed below without asking for further permission.
The arguments to this command are the task. If there are no arguments, ask the user what the task is before doing anything else.
Run the task through the phases, reading each workflow's result before choosing the next step.
fable-understand with args: { focus: "<the subsystem this task touches>" }.fable-design with args: { question: "<the design decision>" } and adopt the synthesized design. If the approach is genuinely forced, write a three-line plan inline instead.fable-migrate with args: { instruction: "..." }.fable-review over the change set (args: { target: "the uncommitted changes", votes: 3 }). Fix every confirmed finding, re-run the tests, and re-run the review if the fixes were non-trivial. The task is done when review comes back clean and tests pass — or when the only remaining findings are ones you explicitly present to the user as accepted risks.fable-ship and hand the deploy trigger to the user.Scale honestly: if the task turns out to be trivial once understood, say so and skip ahead rather than burning a full pipeline on a one-line fix.
If the Workflow tool is not available in this environment, emulate each workflow per the fallback in .claude/fable/FABLE.md.
Loop-until-dry exhaustive hunt — waves of diverse finders keep sweeping until two consecutive rounds surface nothing new, and every fresh finding faces a three-lens judge panel. Use for "find ALL the bugs", full audits, and edge-case sweeps.
Continuous goal-directed operation across many cycles or sessions — maintains a persistent run file (goal, backlog, journal), executes one backlog item per cycle through the full phase discipline, commits at verified milestones, and keeps cycling. Composable with /loop or scheduled tasks for unattended long runs. Use for "keep working on this", overnight runs, or multi-day builds. Wants a git repository — checkpoint commits are its safety rail; it offers git init once if missing.
Repo-wide mechanical change — four parallel discovery modalities find every site, a two-file pilot proves the instruction before the fan-out, then each file is transformed and independently checked, and the project's own checks run once. Use for "rename X everywhere", "replace API A with B", "apply this fix across the repo". It edits files.
Judge-panel design — three genuinely different approaches (simplicity-first, robustness-first, evolution-first) scored by independent judges and synthesized into one grounded implementation plan. Use for architecture decisions or "how should I build X".
Deep codebase research — five parallel search modalities (names, content, structure, git history, tests), deep-reads of every lead, a cited synthesis, and a completeness critic. Use for "where is X handled", "how does Y work", "what breaks if Z changes". This searches the repo, not the web.
Multi-agent adversarial code review — four parallel finders (correctness, contracts, security, resources), with every finding independently attacked by skeptics before it reaches you. Use for "review this", "check my changes", or as a pre-PR gate.