A small collection of evidence-aware, risk-adaptive workflows for AI agents.
This repository explores how to turn an agent workflow into a reusable Skill without sacrificing safety, clarity, or verification.
Turn non-trivial product, document, or software-engineering work into scoped, safe, and verifiable delivery.
It routes work by size and risk:
- small requests use a minimal confirm → act → verify loop;
- complete products and documents use a product-delivery reference;
- engineering work uses an engineering-workflow reference;
- code changes add LLM coding guardrails.
Its central principle is not “make the agent cautious at all times.” It is make safety boundaries explicit while preserving ordinary agency: protect existing work, validate honestly, and require authorization immediately before external side effects.
Apply a portable collaboration policy to multi-step agent work. It makes the operating agreement explicit: act on clear requests, retain already-given choices and authorization, keep changes minimal, validate in proportion to impact, and distinguish evidence from assumptions.
This Skill packages the original Chinese “全局协作规则” as a reusable prompt. It is suited to GPT-6 Astra because stronger instruction following benefits from clear scope and authorization boundaries. It is not an OpenAI-official prompt, policy, endorsement, or guarantee of behavior across every host.
skills/
├── engineering-delivery/
│ ├── SKILL.md
│ ├── agents/openai.yaml
│ └── references/
│ ├── product-delivery.md
│ ├── engineering-workflow.md
│ └── llm-coding-guardrails.md
└── global-collaboration-rules/
├── SKILL.md
└── agents/openai.yaml
This package follows the standard folder-based Agent Skill layout. Copy a skill directory from skills/ into a skills directory recognized by your Agent host, then invoke it explicitly (for example, $engineering-delivery or $global-collaboration-rules) or use a matching request where the host supports automatic skill selection.
Host-specific installation and cross-Agent behavior have not been verified in v0.2. Treat this repository as directly inspectable workflow packages, not a universal compatibility claim.
- The Skills are derived from a personal global collaboration policy, with user profile, machine paths, and private project details removed.
- They do not provide security certification, legal advice, automatic deployment, or universal compatibility guarantees.
- They require explicit user authorization before deployment, publishing, messages, PR creation, deletion, payment, or other external side effects.
- No license has been selected yet; please do not assume reuse rights beyond GitHub's default terms.
v0.2 — structurally packaged and locally validated.
The packages have passed structural Skill validation. That confirms their frontmatter, layout, and absence of unfinished scaffolding; it does not prove real-world behavior, cross-Agent portability, or security completeness.
- Add realistic positive, negative-trigger, and authorization-boundary evaluation cases.
- Run the Skills against representative product and code-change tasks.
- Record observed failures and refine only the rules supported by evidence.
- Verify installation and behavior on additional Agent hosts before claiming portability.
- Choose a license deliberately if the project is intended for reuse.
Maintained by AE0506. Feedback and critical discussion are welcome.