360 Secure SkillHub
360 安全技能中心

Super Spec

核心功能:将零散上下文压缩后生成含需求、技术设计、测试、构建计划等模块的可执行规范,附带溯源矩阵、风险验证路径,解决交接模糊问题。典型场景:需严谨交付物的任务,如产品需求定义、技术方案设计等。输入为上下文文件,输出结构化规范文档,支持Pro模式及 fallback配置,确保可复用。

h-mascot
h-mascot
热度 23.92
最近更新 2026-07-01 02:21:15
分类:效率依赖:代码执行 / 读写文件 / 第三方 API / 内部调用大模型应用场景:报告撰写 / 摘要总结 / 任务管理 / 模板

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add superada-skill-super-spec
Metadata
name
super-spec
description
Create product, technical, PRD, RFC, agent-build, and handoff specs from rough context.

Super Spec

Super Spec turns rough operator context into an executable spec that a builder, reviewer, or agent fleet can act on after context compaction.

Use it when a task needs more than a plan: product requirements, technical design, proof gates, rollout, rollback, builder prompt, and reviewer prompt in one durable artifact.

What it produces

  • Source map: evidence, constraints, assumptions, ignored context, and open questions.
  • Goals and non-goals strong enough to stop scope creep.
  • Requirements with acceptance criteria and validation method.
  • Technical design, data/API surfaces, permissions, dependencies, and failure modes.
  • Tests, evals, proof requirements, production QA gate, rollout, observability, and rollback.
  • Traceability matrix from requirement to proof.
  • Atomic build plan.
  • Builder prompt and reviewer prompt.
  • Route receipt showing whether the run used Pro or an explicit fallback.

Why it exists

Agent work often fails because the handoff is mushy: unclear requirements, hidden assumptions, no proof gate, and no way for another agent to resume after context compaction. Super Spec is the antidote. It compiles messy context into a document that can be built, reviewed, and audited.

Runner behavior

The included runner prefers a Pro/Oracle route when configured. It is fail-closed by default: if Pro is expected and the Pro route fails, it stops instead of silently writing a fallback spec. Use --no-pro for an intentional non-Pro run, or set SUPER_SPEC_ALLOW_CODEX_FALLBACK=on only when a fallback spec is acceptable.

For long Pro jobs, use background Responses mode and enough time:

export SUPER_SPEC_ORACLE_BACKGROUND_MODE=on
export SUPER_SPEC_PRO_TIMEOUT=3600
export SUPER_SPEC_PRO_HTTP_TIMEOUT=60m
export SUPER_SPEC_ORACLE_WALL_TIMEOUT=3700s

Install

curl -sSf https://superada.ai/install/super-spec | sh

Run

super-spec --title "Enterprise Local Model Orchestrator" --input context.md

Or directly:

skills/super-spec/scripts/run-super-spec.sh --title "<title>" --input <context-file>

Configuration

Set these when using a Pro route:

export SUPER_SPEC_PRO_MODEL=gpt-5.4-pro
export SUPER_SPEC_PRO_BASE_URL=http://127.0.0.1:4000/v1
export SUPER_SPEC_PRO_API_KEY=...

For Azure Responses via Oracle, set your own deployment values:

export SUPER_SPEC_AZURE_ENDPOINT=https://your-resource.openai.azure.com
export SUPER_SPEC_AZURE_DEPLOYMENT=gpt-5.4-pro
export SUPER_SPEC_AZURE_API_VERSION=2025-04-01-preview

Do not paste API keys into specs, prompts, or public docs.

Quality bar

  • The spec must reduce degrees of freedom for the builder.
  • Requirements must be testable or explicitly unresolved.
  • Non-goals must block obvious scope creep.
  • Every risky assumption needs a validation path.
  • Tasks must be small enough for review and proof.
  • The final spec must be resumable by another agent after context compaction.