English | 简体中文
PromethOS is a skills-first framework for designing, governing, evaluating, and evolving agent systems across domains.
It packages reusable SKILL.md prompts for agent requirements, architecture,
context, memory, permissions, tool contracts, guardrails, evaluation, and
library maintenance.
PromethOS is currently in an early public preview: the shared skills/
contract is stable enough to use and extend, while installation and packaging
remain intentionally lightweight through the skills CLI and manual fallback
docs.
PromethOS is downstream of Claude Code agent-design practice, but it is not a copy of the Claude Code runtime.
The library extracts reusable agent-development principles from upstream runtime experience, then expresses them as portable skills, references, and evals for broader reuse.
Maintainer-facing abstraction docs live under docs/framework/README.md.
PromethOS gives an agent a structured way to answer:
- what this agent system is supposed to do,
- how its subsystems should be designed well,
- where autonomy and risk boundaries should be drawn,
- how the agent should be evaluated and improved,
- how the skills library itself should evolve.
The library is intentionally lightweight: one canonical skills/ directory,
plus thin platform-specific entrypoints for the harness you are using.
When PromethOS coexists with a broader workflow library, PromethOS is intended
to act as the agent-domain methodology layer rather than a replacement for all
session-level process control.
-
For Codex, install PromethOS globally with the
skillsCLI:npx skills add bikeread/promethos -g -a codex -s '*' --copy -y -
Restart Codex, then verify that the bootstrap routing skill is visible.
-
If you prefer manual install or use another harness, follow the install guide for Codex, Claude Code, or Gemini CLI.
- A routing/bootstrap skill decides whether PromethOS should lead and which agent-design question comes first.
- Core entry skills handle the first high-leverage decisions: requirements, autonomy boundaries, eval design, debugging, and readiness checks.
- Design-deepening skills handle architecture, implementation planning, context, memory, permissions, tool contracts, delegation, and scope.
- Maintainer skills keep the library itself coherent and improve it from evidence.
| Platform | Status | Install Guide |
|---|---|---|
| Claude Code | Documented | docs/README.claude-code.md |
| Codex | Documented | .codex/INSTALL.md / docs/README.codex.md |
| Gemini CLI | Compatible | docs/README.gemini.md |
PromethOS is being organized around three public layers:
- Core entry:
route-agent-design,define-agent-requirements,set-agent-autonomy-boundaries,build-agent-evals,debug-agent-failures, andverify-agent-readiness - Design deepening: architecture, implementation, context, memory, permissions, tools, delegation, and scope control
- Maintainer workflows: library evolution and turning incidents into reusable improvements
PromethOS is designed to work both alone and alongside broader workflow skill libraries such as superpowers.
- If PromethOS is installed by itself, the routing/bootstrap skill can bootstrap the full methodology flow.
- If a stronger session-level workflow system is already active, let it own generic process control such as brainstorming, generic implementation planning, and session-wide execution discipline.
- In coexistence mode, use PromethOS for agent-system requirements, architecture, context, memory, permissions, autonomy, evaluation, and library evolution.
See docs/COMPATIBILITY.md for the coexistence rules.
skills/
docs/
.codex/
.github/
GEMINI.md
gemini-extension.json
scripts/
Key maintainer docs:
Contributions are welcome. Start with CONTRIBUTING.md before adding or restructuring skills. For public vulnerability reports, follow SECURITY.md instead of opening a detailed public issue.
- Community link: LINUX DO
PromethOS is released under the MIT License. See LICENSE.