polyak-dev-kit

A personal set of Claude Code commands, agents, and skills, packaged as an installable plugin.

Install

/plugin marketplace add jpol34/polyak-dev-kit
/plugin install toolkit@polyak-dev-kit

Claude Code will prompt for one config option at install time — see Configuration below.

What's included

Commands

• ss — grab the image currently on the clipboard and look at it. Windows-only (uses System.Windows.Forms.Clipboard via PowerShell); no macOS/Linux equivalent is included.
• monitor — set up event-filtered monitoring for a running background process or subagent so progress/failure surfaces without manual polling.
• explain — a concise, plain-English explanation of a topic, file, or plan.
• plan-review — send the current plan to the plan-critic agent for adversarial review before building.

Agents

• plan-critic — reviews someone else's plan: verifies its claims against the actual codebase, attacks its riskiest assumptions, and ends with a ship/revise/different-approach verdict.

Skills

• grilling (and the grill shorthand) — a relentless, round-based requirements interview that maps decisions as a design tree and proposes a recommended answer for each question.
• research — turns a loose topic into a sourced findings report via parallel research agents.
• to-tickets — breaks a plan into dependency-ordered, independently-buildable tickets and publishes them to GitHub Issues (or local files).
• frontend — detects a repo's current frontend state and brings it to a cohesive, professional look via a component-system install/retrofit plus a screenshot-critique loop.

Docs

• docs/CLAUDE.md — the actual workflow philosophy these tools are built around (the discuss→plan→review→ticket→build→ship pipeline, comment/doc discipline, token discipline, secrets handling), included as a reference to adapt for your own CLAUDE.md, not something this plugin loads automatically.

Not included (kept local, not published here): a full PR-through-deploy pipeline command and two agents that verify a live deploy / run a local test suite — those are tied to a specific deploy platform and aren't generalized yet. to-tickets step 6 hands off to "run your own PR-through-merge/deploy process" rather than naming a command that isn't here.

Configuration

This plugin declares one userConfig option, prompted at install/enable time:

• auto_continue (boolean, default false) — when true, plan-review auto-continues into to-tickets once a plan is settled instead of stopping to ask, and to-tickets skips the merge-confirmation gate after the first ("canary") ticket in a build run instead of confirming every ticket. Leave this off until you trust the pipeline in your own repos — it changes what gets merged without asking. Change it later with /plugin configure toolkit@polyak-dev-kit.

Manual setup: statusLine

One hook script, context-statusline.cjs, can't be wired up automatically — Claude Code's statusLine setting only lives in your own ~/.claude/settings.json, and plugins can't declare it. The plugin bundles the script; you point your own statusLine at it once, after install:

1. After installing, ask Claude to copy the script out of the plugin's install location to a stable path: ${CLAUDE_PLUGIN_ROOT} resolves inside plugin/hook content but not inside a manually-edited settings.json, and the plugin's on-disk install path is version-pinned (.../plugins/cache/polyak-dev-kit/toolkit/<version>/...), so pointing statusLine directly at it would break on every plugin update. Copy hooks/scripts/context-statusline.cjs from the installed plugin to ~/.claude/hooks/context-statusline.cjs (a location outside the plugin cache that survives updates).
2. Add this to your own ~/.claude/settings.json:

   "statusLine": {
     "type": "command",
     "command": "node \"$HOME/.claude/hooks/context-statusline.cjs\""
   }

The other hook, context-compact-check.cjs, installs and runs automatically — no manual step needed for it.

Dependencies this plugin assumes

Not enforced by the manifest — Claude Code has no mechanism to declare "requires an externally-configured MCP server," so noting it here instead:

• to-tickets assumes gh (CLI) or a github MCP server for its default GitHub Issues tracker; it falls back to local files if you tell it to use a different tracker.
• plan-critic uses WebFetch/WebSearch to verify claims when relevant — no special setup needed beyond what's built into Claude Code.
• frontend checks for and offers to install Anthropic's own frontend-design@claude-plugins-official plugin for aesthetic guardrails; it can proceed without it if you explicitly say to.
• frontend's visual verification loop uses a Playwright or chrome-devtools MCP server if one is configured; it degrades to a static code review if neither is available.

Known behavior: installing alongside your own commands of the same name

Verified by installing this plugin into a profile that already had its own ~/.claude/commands/explain.md: the plugin's files install into a separate, version-pinned cache directory (~/.claude/plugins/cache/polyak-dev-kit/toolkit/<version>/...) rather than overwriting anything in ~/.claude/commands/. Both files coexist on disk. Invocation-time behavior (which one actually runs if you type the bare command name) wasn't verified beyond that — if you already have your own command with the same name as one in this plugin, expect to need to check which one wins or rename yours.

License

MIT — see LICENSE.