Runewright

A skill and subagent factory for LLM agents. One GitHub marketplace supports two runtimes: Claude Code and Codex. Runewright turns a workflow into a small, source-grounded route that strong and smaller models can execute consistently.

Included skills

Skill	Purpose
skill-brief	Runs a one-to-three-round interview and records the task, triggers, sources, criteria, boundaries, and gold example.
skill-generator	Turns a request into a production-ready SKILL.md; automatically runs skill-brief when no ready brief exists.
subagent-generator	Creates a Claude Code Markdown role and/or Codex TOML role; automatically runs skill-brief when needed.
skill-audit	Audits environment skills for duplication, bloat, inactivity, and blueprint compliance, then writes a remediation plan.
runewright-blueprint	Defines the seven required elements, anti-patterns, language policy, and live-skill feedback contract used by the other skills.

Typical loop: skill-brief -> skill-generator -> use the skill -> run skill-audit periodically. Every generated skill keeps an append-only execution log and can incorporate recurring feedback during later consolidation.

Language behavior

The repository, manifests, templates, and skill instructions are English. At runtime, Runewright:

• speaks in the language of the user's current request;
• writes a generated brief, skill, or agent in an explicitly requested language;
• otherwise writes the artifact in the language of the user's current request;
• keeps code, identifiers, paths, source IDs, YAML/TOML keys, and status tokens such as DONE, PARTIAL, and BLOCKED stable.

Install from the GitHub marketplace

Claude Code

From a terminal:

claude plugin marketplace add TinyFrontier/runewright
claude plugin install runewright@runewright

Or from an interactive Claude Code session:

/plugin marketplace add TinyFrontier/runewright
/plugin install runewright@runewright

Invoke skills explicitly as /runewright:skill-brief, /runewright:skill-generator, /runewright:subagent-generator, and so on. Claude can also select them from their descriptions. Calling either generator without a ready brief starts the brief interview automatically unless the user explicitly asks to skip it.

Codex

codex plugin marketplace add TinyFrontier/runewright
codex plugin add runewright@runewright

Start a new task after installation so Codex loads the plugin components. Invoke skills explicitly as $runewright:skill-brief, $runewright:skill-generator, and so on, or let Codex select one from its description.

Both marketplace commands can use a private repository when local Git is already authorized. For public installation, make this repository public and publish the files from the release branch.

Local Codex development without a marketplace

git clone https://github.com/TinyFrontier/runewright.git
cd runewright
./install-codex.sh
./install-codex.sh --project /path/to/repo

The script creates symlinks in ~/.agents/skills or <repo>/.agents/skills, so git pull updates locally installed skills without a plugin reinstall. If a destination is an ordinary file or directory, the script reports a conflict and does not overwrite it.

Project state

Generated and accumulated state lives in the platform-neutral .runewright/ directory at the project root:

.runewright/
  sources/    # truth sources: INDEX.md, decisions/, external/
  feedback/   # skill and agent execution logs
  briefs/     # briefs written by skill-brief
  audits/     # skill-audit reports

skill-generator copies the initial sources/ scaffold from skills/skill-generator/assets/sources-scaffold/ on first use.

Design decisions

• A skill is a route, not a textbook. Knowledge lives under .runewright/sources/: derived truth is represented by a retrieval command, decided truth by a decision file with rationale, and external truth by a pointer with its last verification date. Skills cite stable IDs from a flat INDEX.md.
• One skill format serves both platforms. Agent Skills frontmatter contains name and description; the workflow lives in the body. Relative skill links, such as ../runewright-blueprint/SKILL.md, work in Claude's plugin cache and through Codex skill symlinks.
• Agent roles use two native formats. subagent-generator writes YAML-frontmatter Markdown for Claude Code and TOML with developer_instructions for Codex while keeping one role contract and report format.
• Hard size limits. Skill body <= 150 lines, agent body <= 100 lines, and description <= 500 characters. Short descriptions reduce persistent context load and remain useful when a client truncates long skill lists.
• No automatic deletion. skill-audit proposes a plan. Destructive changes require separate user confirmation.

Release validation

claude plugin validate . --strict
bash -n install-codex.sh

Update the release version together in .claude-plugin/plugin.json and .codex-plugin/plugin.json. Current version: 0.3.0.

Limitations

• install-codex.sh requires a POSIX shell and is only needed for direct skill installation. Windows users should prefer marketplace installation.
• Claude Code and Codex use different subagent configuration formats, so a cross-platform role is represented by two files.