A catalog of plugins and reference skills for AI coding agents — Claude Code, OpenAI Codex, Google Antigravity (AGY), and Pi. The plugins cover requirements-first delivery, adaptive opportunity scanning, grounded research, UI mockups, code audits, prose craftsmanship, and cross-agent coordination.
The centerpiece is Workbench. You describe an outcome in plain language; the agent learns the repository, asks you for the decisions that belong to you, and drives the agreed scope to a verified finish. Decisions, evidence, and delivery state live in the repository, not in the chat that produced them.
You need Claude Code with Task/Agent tool support, the OpenAI Codex CLI,
Google Antigravity (agy), or Pi. To adopt Workbench in a project, that project must be a Git repository.
Pick your harness below. Install Workbench first, then add any standalone plugins you want.
/plugin marketplace add nklisch/skills
/plugin install workbench@nklisch-skills # the centerpiece
/plugin install ux-ui-design@nklisch-skills # mockup-first UI design
/plugin install code-audit@nklisch-skills # markdown code audits
/plugin install nates-toolkit@nklisch-skills # utility skills
/plugin install agentic-research@nklisch-skills # grounded research discipline
/plugin install agent-coordination@nklisch-skills # cross-agent ledger
/plugin install prose-craft@nklisch-skills # prose drafting and review
/plugin install theatre-feedback@nklisch-skills # Theatre operating skills and feedback notices
/plugin install declaudify@nklisch-skills # every-turn writing posturecodex plugin marketplace add https://github.com/nklisch/skills
codex plugin install workbench
codex plugin install ux-ui-design
codex plugin install code-audit
codex plugin install nates-toolkit
codex plugin install agentic-research
codex plugin install agent-coordination
codex plugin install prose-craft
codex plugin install theatre-feedbacktheatre-feedback adds Stage and Director operating guidance and an optional
pending-feedback hook. Install the Theatre CLI and Godot addons separately from
nklisch/theatre; the plugin does not install
or register the MCP servers. Hook trust remains an explicit client choice. The
current 0.3.4 prebuilt Theatre release predates the feedback-hook helper, so
build Theatre from current source if you want pending-feedback notices before a
newer release is available. The operating skills remain useful without the hook.
Clone or vendor the repository and add the desired plugins to your project (.agents/plugins.json) or global (~/.gemini/config/plugins.json) configuration:
{
"entries": [
{ "path": "path/to/skills/plugins/workbench" },
{ "path": "path/to/skills/plugins/ux-ui-design" },
{ "path": "path/to/skills/plugins/code-audit" },
{ "path": "path/to/skills/plugins/nates-toolkit" },
{ "path": "path/to/skills/plugins/agentic-research" },
{ "path": "path/to/skills/plugins/agent-coordination" },
{ "path": "path/to/skills/plugins/prose-craft" }
]
}Pi installs go through pi-plugins,
the plugin marketplace and lifecycle manager. This repo no longer publishes
npm packages — pi-plugins does.
pi install npm:@nklisch/pi-pluginsThen, inside Pi:
/plugins marketplace add nklisch/skills
/plugins add workbench@nklisch-skills --scope user
/plugins add ux-ui-design@nklisch-skills --scope user
/plugins add code-audit@nklisch-skills --scope user
/plugins add nates-toolkit@nklisch-skills --scope user
/plugins add agentic-research@nklisch-skills --scope user
/plugins add agent-coordination@nklisch-skills --scope user
/plugins add prose-craft@nklisch-skills --scope user
/plugins add theatre-feedback@nklisch-skills --scope user
Use --scope project instead of --scope user to enable a plugin for one
repository only.
The Pi-native tool packages publish to npm from the pi-extensions repository:
pi install npm:@nklisch/pi-background-tasks # background jobs and monitors
pi install npm:@nklisch/pi-zai-research # Z.ai research toolspeeragent (cross-harness peer delegation for implementation, research, and review) is an external companion. Install it through the bridge or from its own repository:
/plugins add peeragent@nklisch-skills --scope user
pi install git:github.com/nklisch/peeragent@v0.6.0After installing, ask your agent:
Set up Workbench in this repository.
Workbench setup proposes a repository-grounded working agreement for collaboration,
review, verification, and completion. Accept it together or adjust the choices;
setup then writes the confirmed conventions and a small .work/ ledger.
It also reconciles project conventions, creates an empty canonical pattern
index for future evidence-backed extraction, can offer CLAUDE.md as a symlink
to canonical AGENTS.md, and projects the pattern catalog into Claude's skill
directory when applicable. During large runs, useful recurring evidence can justify
an extraction proposal, but only a selected outcome authorizes that work.
From there, direct the agent in ordinary language. work owns the continuous
outcome, using design and deliver guidance without repeated handoffs.
Quick implementation and focused review often stay in the current context;
another agent earns its place through useful expertise, independence, or throughput.
deliver also remains a direct entry point for one named ready item.
For example: "Implement this outcome,"
"Deliver the ready rate-limiting feature," "Drive the onboarding epic to done,"
or "Research the prior art for this decision."
The Workbench guide walks through adopting a repo and driving work; the plugin README covers the mental model — autonomy settings, review weights, design lenses, and the research discipline.
Workbench and agile-workflow both own a .work/ directory, and they are
mutually exclusive within one project. Pick one:
| Pick | Best fit | Status |
|---|---|---|
| Workbench | Requirements-first delivery through ordinary conversation, with a compact ledger and integrated research. Consolidates older workflows — including an existing agile-workflow substrate — through setup. |
Active development. Recommended for new projects. |
| agile-workflow | Projects that already rely on its explicit stage-and-gate pipeline and autopilot queue runner. | Supported in maintenance mode: bug fixes and compatibility work, no new feature development. |
Deprecated: the older
workflowplugin (doc-driven,docs/designs/) is unmaintained. Existing installs keep working. Migrate to agile-workflow with/agile-workflow:convert, or consolidate onto Workbench with/workbench:setup. The old guide remains at docs/workflow-guide.md for reference.
| Plugin | What it does | More |
|---|---|---|
| workbench | Continuous, requirements-first delivery with a compact .work/ ledger, proportionate design and review, and grounded .research/ evidence when needed. |
Guide · README |
| ux-ui-design | Mockup-first UI design. Throwaway single-file HTML mockups in .mockups/ to align on direction before production code. |
docs/ux-ui-design-guide.md |
| code-audit | Markdown-first audits with no workflow dependency: deep code, bug, security, and test scans; perf scouting; repo scorecards. | plugins/code-audit/README.md |
| nates-toolkit | Project-agnostic utilities — plain-language re-explanation, agent self-reflection, skill authoring and auditing. | plugins/nates-toolkit/README.md |
| agentic-research | Grounded, verifiable research discipline with attestations and citation gates. Stands alone or pairs with a work tracker. | plugins/agentic-research/README.md |
| agent-coordination | Sparse cross-agent coordination ledger for shared repositories, backed by GitHub Discussions. | plugins/agent-coordination/README.md |
| prose-craft | Prose craftsmanship for human-facing docs: style-contract drafting, six-lens review, multi-model refine cycle. | plugins/prose-craft/README.md |
| theatre-feedback | External Theatre companion with Stage and Director operating skills plus optional pending Godot feedback notices. Theatre itself is installed separately. | nklisch/theatre |
| declaudify | Claude Code-only every-turn writing guidance: simple, deeply informative technical language with framed codebase references and no invented shorthand. | plugins/declaudify/README.md |
| agile-workflow | Structured stage-and-gate work tracking with a goal-backed autopilot queue. Supported in maintenance mode — see above. | docs/agile-workflow-guide.md |
Beyond the plugins, .agents/skills/ carries a curated library of reference
skills that auto-load when their library or topic appears in the work — Bun,
Zod v4, Biome v2, Drizzle, Hono, the TanStack family, Zustand, citty,
@clack/prompts, smol-toml, schemars, and others. A few standalone utilities
(clean-memory, design-pages) and ecosystem research notes round it out.
plugins/<name>/ # one directory per plugin; shared skills/
# plus channel metadata (declaudify is Claude-only)
.agents/skills/ # standalone reference-skill library (non-plugin)
.claude-plugin/ # Claude Code marketplace catalog
.agents/plugins/ # Codex marketplace catalog
scripts/bump-version.sh # per-plugin version gate across all channel metadata
docs/ # repo-level vision, spec, architecture, and guides
Every cross-channel supported plugin ships the same skills/ directory to all
three native harnesses and keeps its Claude and Codex manifests in version
lockstep. Channel-specific plugins may intentionally expose only the host
surface they target; declaudify is authored for Claude Code and is absent from
the Codex and Antigravity catalogs. Pi's bridge can discover Claude catalog
entries, but Pi support is not provided or claimed for this plugin.
The distribution rules live in docs/SPEC.md; the layout and
wiring live in docs/ARCHITECTURE.md; the purpose and
dogfooding thesis live in docs/VISION.md.