Open Spec · SKILL.md

A Skill is a portable AI capability, not just a prompt.

Prompts are disposable. Skills are durable. A Skill (defined in a SKILL.md file) declares what an AI capability takes as input, what it produces as output, how it executes, and what governance rules it must follow — then carries all of that into any agent runtime, unchanged.

A SKILL.md file is a contract, not a conversation.

When you paste a System Operating Procedure into Cromus, the SOP Compiler turns it into a structured SKILL.md file. That file is the skill. It travels anywhere — into Claude Code, LangGraph, CrewAI, n8n, or any MCP-compatible agent — and it brings its governance rules with it.

A skill is also the unit of measurement in Cromus. Every skill is scored for cost, latency risk, failure risk, and structural gaps — before any execution. The waste metric (Croms™) lives at the skill level.

The SKILL.md format is an open specification. You can read it, edit it by hand, validate it in the free Cromus validator, and store it in version control like any other file.

For teams whose SOP documents live in a repo, see the CI compiler desk.

content-review.skill.md
name: content-review-v2
description: Review marketing copy for compliance and tone
inputs:
- draft_copy (required)
- tone_profile (optional)
outputs:
- review_report.md
- compliance_score
steps:
1. Extract claims from draft
2. Check compliance rules
3. Score tone against profile
4. Generate structured report
governance: → ETHOS.md
max_cost_per_run: $0.05
quality_tier: balanced

Anatomy of a skill

Every field in a SKILL.md has a purpose. Nothing is decorative — every declaration is used by the compiler, the scorer, or the runtime.

name:
Name

A unique identifier for the skill — machine-readable and human-readable.

e.g. content-review-v2
description:
Description

What the skill does in one plain sentence. Written for the runtime, not the author.

e.g. Review marketing copy for compliance and brand tone.
inputs:
Inputs

Declared required and optional inputs with types. No ambiguity about what the skill needs.

e.g. draft_copy (required), tone_profile (optional)
outputs:
Outputs

Named output artifacts the skill produces. Makes integration explicit.

e.g. review_report.md, compliance_score
steps:
Steps

Ordered, named execution steps. Each step is scoreable and optimizable independently.

e.g. 1. Extract claims · 2. Check compliance · 3. Score tone · 4. Generate report
governance:
Governance (ETHOS.md)

Behavioral rules, budget ceilings, and policy constraints that travel with the skill.

e.g. max_cost_per_run: $0.05 · disallow_models: [gpt-5.4-pro]

Skill vs. prompt

Prompts work for one conversation. Skills work for production.

SKILL.md
Plain prompt
Versioned
—
Portable across runtimes
—
Has declared inputs & outputs
—
Carries governance rules
—
Cost-scoreable before execution
—
Human-readable
Quick to write
—

What you can do with skills

Skills are the unit of work in Cromus. Here's what becomes possible once your AI capabilities live in structured files.

Replace ad-hoc prompts

Every one-off prompt your team copies into Slack is a skill waiting to happen. Skills make that knowledge durable, versioned, and reusable.

Enforce policy before execution

Governance rules live inside the skill file — budget ceilings, disallowed models, and tone constraints travel with the artifact into any runtime.

Version and compare

Skills are files. They live in version control. You can diff them, roll them back, and run A/B experiments on skill versions using real run data.

Build composable workflows

Complex workflows are composed from skills. Each skill has declared outputs that become another skill's inputs — no hidden coupling.

Distribute as packs

Skill packs are collections of pre-built skills for common use cases — content review, lead scoring, SEO research, and more. One-click install.

Score cost before you pay it

The Croms engine simulates every skill's cost against verified model pricing before you make a single LLM inference call. Waste is visible before it happens.

Open Spec Stack

Skills don't travel alone.

The SKILL.md spec is one of five open specifications in the Cromus stack. Each covers a different layer of AI workflow governance:

SKILL.md
Capability definition

What the AI does — inputs, outputs, steps, model requirements, and structural metadata. The execution contract.

ETHOS.md
Behavioral governance

What the AI is allowed to do — budget ceilings, disallowed models, tone constraints, and escalation rules. Travels with the skill.

MEMORY.md
Portable context

What the AI remembers between runs — durable lessons, decisions, and institutional knowledge that persist across sessions and runtimes.

GRAPH.md
Workflow topology

How capabilities connect — nodes, edges, routes, bounded loops, shared state, and path budgets across the whole workflow.

HARNESS.md
AI system manifest

How prompts, guardrails, memory, tools, subagents, skills, scripts, and plugins fit together as one governed AI system.

One skill file. Every runtime.

Cromus exports skills in formats that any modern agent runtime can consume. You write the capability once; the export target handles the wrapper.

Claude Code
Replit Agent
LangGraph
CrewAI
n8n
AutoGen
Ollama
Open WebUI
Codex
Cowork

Remote MCP Server

Use skills from inside your AI client

Connect the Cromus MCP server to Claude or any MCP-compatible client. Score and validate skills without leaving the conversation.

Connect

Your first skill takes 60 seconds.

Paste any SOP, workflow description, or procedure into the Cromus compiler. It produces a SKILL.md file with a cost score, Croms waste rating, and an export-ready package — no account required.