Agent Skills¶

Karenina ships a set of agent skills — structured instruction packs for coding agents (Claude Code and compatible harnesses) — that encode the correct karenina workflows: benchmark construction, template and rubric authoring, pipeline configuration, result analysis, and the CLI. They live in skills/ at the repository root and are versioned together with the library, so they always match the API of the commit you have checked out.

What is included¶

Skill	Purpose
`using-karenina`	Router: picks the right workflow skill from a request; carries cross-cutting gotchas and the full reference docs under `references/`
`karenina-qa`	Single-turn QA benchmarks: questions, templates, checkpoints, runs
`karenina-scenarios`	Multi-turn branching scenario graphs, outcome criteria, TurnChecks
`karenina-task-eval`	Scoring pre-recorded outputs (logs, traces, JSON exports)
`karenina-template-authoring`	`BaseAnswer` / `VerifiedField` templates and verification primitives
`karenina-rubric-authoring`	Rubric traits (LLM, regex, callable, metric, agentic) and `DynamicRubric`
`karenina-verification`	`VerificationConfig`, adapters, guards, presets, resume
`karenina-results`	Loading, filtering, DataFrames, comparison, export, repair
`karenina-cli`	Terminal workflows: `karenina verify`, presets, progressive save/resume
`karenina-manual`	Full pipeline with manually supplied answers (`ManualAdapter`)
`testing-gate`	Pre-merge test battery: offline unit gate + opt-in live tests
`karenina-adapter-*`	Building new karenina adapters (gather → design → implement → test → review)

Installation¶

From a karenina checkout, the installer copies (or symlinks) the skills into the directory your agent scans:

./skills/install.sh                      # Claude Code, this project
./skills/install.sh --agent codex        # Codex, Pi, Gemini CLI, OpenCode, ...
./skills/install.sh --scope user         # available in every project
./skills/install.sh --workflow-only      # skip the adapter development skills
./skills/install.sh --help               # all options

To install by hand, copy the skill directories yourself:

Agent	Project scope	User scope
Claude Code	`.claude/skills/`	`~/.claude/skills/`
Codex	`.agents/skills/`	`~/.agents/skills/` (also reads `~/.codex/skills/`)
Pi	`.agents/skills/`	`~/.agents/skills/`
Gemini CLI, OpenCode, Goose, Amp, ...	`.agents/skills/`	`~/.agents/skills/`

mkdir -p .agents/skills
cp -r /path/to/karenina/skills/karenina-* /path/to/karenina/skills/using-karenina .agents/skills/

.agents/skills/ is the shared convention that most agents now scan, so skills installed there are visible to all of them at once.

Working inside this repository there is nothing to do: .claude/skills is a symlink to skills/, so Claude Code discovers them automatically.

Copy only the skills you need. Every skill is self-contained, except that the workflow skills refer to using-karenina/references/ for deep-dive documentation, so keep using-karenina alongside them.

Keeping the reference docs in sync¶

skills/using-karenina/references/ is generated from docs/ — never edit it by hand. After changing the docs, run:

make sync-skill-docs

Using the skills without an agent¶

Each SKILL.md is plain Markdown and reads as a task-oriented guide with verified code snippets, gotcha lists, and API tables. Even without a coding agent they are a practical quick-reference for the workflows they cover.