Skip to main content

orchestrate

Run a feature or fix through an opinionated plan -> spec -> build -> verify cycle with a hard approval gate, instead of hand-rolling subagent orchestration. Use for multi-step work where you want a plan and spec approved before any edit, and every step verified. Trigger with /orchestrate or "orchestrate this", "plan spec build verify", "run this with a gate".

跳到安装

来源信息

仓库
Zavelinski/orchestrate
最近来源活动
2026年6月29日 23:57
检测到的 SKILL.md 语言
英语
星标
0
分支
0

安装方式

默认使用会先检查来源的 Prompt;你也可以切换为直接命令,或下载本地副本。

检查来源文件

决定是否安装前,请先阅读 SKILL.md,以及 SkillsMP 当前展示的配套文件。

正在显示 SKILL.md

SKILL.md
来源说明 · 只读预览
name
orchestrate
description
Run a feature or fix through an opinionated plan -> spec -> build -> verify cycle with a hard approval gate, instead of hand-rolling subagent orchestration. Use for multi-step work where you want a plan and spec approved before any edit, and every step verified. Trigger with /orchestrate or "orchestrate this", "plan spec build verify", "run this with a gate".
version
0.1.0
user-invocable
true
metadata
{"emoji":"🎼"}
# orchestrate A packaged orchestration cycle so you stop hand-rolling planner/spec/builder/verifier every time. Built on the Workflow tool, with a hard gate (no edits before approval) and a `verify:` check attached to every step. ## Why this exists (evidence) - Power users say agent orchestration (planner/spec/builder/verifier) is DIY, and that agents change too much code with weak gating. This packages the pattern. - Orchestration moves the needle: AI21 Maestro reaches top-tier SWE-bench Verified purely by allocating compute across plan + multi-path execution. Self-consistency (sampling + vote) adds up to +17.9pp on reasoning. The cycle below applies both ideas: structured planning plus adversarial verification. ## The cycle ``` FRAME -> restate the goal as a verifiable outcome PLAN -> steps, each with an explicit verify: check SPEC -> a dedicated sub-agent writes the spec to a file GATE -> HARD STOP: present plan + spec for user approval; no edits before OK BUILD -> execute each step, surgically; run its verify: immediately VERIFY -> adversarially confirm the whole change does what was claimed ``` Rules: - The GATE is real. Produce plan + spec and stop. Do not edit files until the user approves. In a headless Workflow there is no interactive prompt, so the gate = phase 1 returns the plan/spec and you (or the user) review before invoking phase 2 (build). - Every PLAN step carries a `verify:` (a test, command, or check). A step with no verify is not done, it is hopeful. - BUILD is surgical: change only what the step requires (compose with karpathy / surgical-diff). - VERIFY is adversarial: try to falsify the claim, do not just re-assert it (compose with adversarial-verify / testsmith). ## How to run - Small task: follow the cycle inline, stop at GATE for approval, then build + verify. - Larger task: use the bundled Workflow template `workflows/plan-build-verify.js` as a starting point. It does the FRAME/PLAN/SPEC fan-out and an adversarial check of the spec, returns the result (that is the gate), and you invoke the build phase after approval. Copy it into your project's `.claude/workflows/` and adapt. - Route cheap stages (drafting, scanning) to a cheaper model via model-router; keep judging/verifying on the strong model. ## Composes with - `writing-plans` / `brainstorming`: front of the cycle. - `karpathy-coding` / `surgical-diff`: keep BUILD minimal. - `adversarial-verify` / `testsmith`: the VERIFY stage. - `model-router`: per-stage model cost. - `context-warden`: keep each stage's context lean. ## Honest limits - The hard gate depends on you honoring it; the skill enforces the discipline, not the runtime. In headless mode the "gate" is a phase boundary, not an OS-level block. - The Maestro/self-consistency numbers are from their benchmarks; the cycle is sound engineering, but measure your own outcome (tests green, scope held).
在 GitHub 查看