A small collection of portable Agent Skills — a capability (a runnable CLI
under scripts/) plus authored guidance (SKILL.md) that AI coding agents such as Claude Code,
Codex, and OpenCode can install and invoke. Each skill is a self-contained folder; the capability
also works from a plain shell, and the skill adds discovery, judgment, and workflow on top.
Everything here is Apache-2.0 licensed. Status varies per skill — check each skill's SKILL.md and
CHANGELOG.md.
| Skill | What it does |
|---|---|
debate |
Run a cross-vendor, recorded, resumable multi-model debate from the shell: propose → blinded critique → revise → (dialectical adversary) → arbitrate, across LLMs from different vendors, with every call written to disk and cached. Reach for it to pressure-test a decision, steelman or critique a claim/paper, or get surfaced disagreement from an inspectable panel. Pre-alpha (v0.2.x): set-generation / steelman debates work; real runs use remote providers and can incur cost. |
software-engineering |
Apply a disciplined, repository-aware engineering workflow to non-trivial code work: resolve local project authority, select the smallest relevant guidance from a bundled knowledge library, establish a baseline before editing, run architectural fitness checks, and bind completion claims to evidence. Ships a deterministic engineering CLI (policy, checks, classification, documentation, inspection) plus a curated knowledge corpus. |
Each skill folder is self-describing — start with its SKILL.md.
These are skills.sh skills, installed with the skills CLI via npx (no global
install required). Check npx skills --help for the current flag set.
# add one skill from this monorepo
npx skills add markov-root/skills --full-depth --skill debate
# install into specific agents (repeatable)
npx skills add markov-root/skills --full-depth --skill software-engineering \
--agent claude-code --agent codex --agent opencodeManage installed skills with npx skills list, npx skills update, and npx skills remove. Prefer
a global symlink install so every supported agent shares one canonical copy. Set DISABLE_TELEMETRY=1
(or DO_NOT_TRACK=1) for a telemetry-free install.
Prerequisites. The code-bearing skills run on Python 3.11+ through
uv — install uv first; the first uncached invocation fetches locked
dependencies. debate additionally needs a provider CLI (claude or codex) or an OpenRouter key
for real runs; offline/fake runs need neither.
If you are an AI agent with shell access, install exactly as above — npx skills add … registers the
skill for your harness, after which its SKILL.md trigger surfaces it to you automatically. You do
not need to read a whole skill up front: match the task to the skill's SKILL.md description, then
invoke it. Re-run npx skills update to pick up new versions.
These skills are plain CLIs and need no MCP server. If you separately run
Model Context Protocol servers, we recommend
mcpm.sh (MCPM) for MCP discovery, installation, profiles, updates, and
per-client configuration — the MCP-side complement to what skills.sh does for skills. Keep the two
concerns separate: skills for capabilities and guidance, mcpm for MCP server provisioning.
Apache-2.0. See each skill's LICENSE and THIRD_PARTY_NOTICES.md.