skill

claude-kaizen

Continuous improvement for Claude Code: measures how you really work, lets you pick fixes in an interactive console, then configures hooks, skills, agents and rules for you. /insights that acts.

KROU4/claude-kaizencommunitylistedunreviewed

Use skill

Set up

Skills are installed by placing their folder where your agent loads skills.

Download ZIP from GitHub Open the repository

The download comes from github.com/KROU4/claude-kaizen itself, not from our server, so what you get is exactly what its author published.

Install instructions come from the README

Skills do not use an MCP config. Follow the steps in the README below.

README

synced Nov 1, 2026

Claude Code's built-in /insights tells you where your sessions go wrong. Then it leaves you to write the hooks, rules and allowlists yourself. Kaizen closes that loop:

measure  →  recommend  →  you choose  →  apply  →  verify  →  (undo)  →  measure again in a week

What it does

Stage	What happens	Where
Measure	Streams every transcript in `~/.claude/projects` (1 GB ≈ 10 s), reads `/insights` facets when present, inventories your settings, hooks, skills, agents, plugins and MCP servers. Produces `measure.json` and a 200-line digest `measure.md`.	`node scripts/kaizen.mjs measure`
Recommend	Evaluates 31 evidence-based recipes in 12 categories against the measurement. Each recommendation carries the exact numbers that triggered it.	`… recommend` (`--add FILE` for extra recipes)
Choose	An interactive console: in the Claude desktop app it is a live artifact Claude reads back; in a terminal it is a local web page; over SSH it falls back to multiple-choice questions.	`… console`, `… choices --import FILE`
Apply	Writes hooks into `settings.json`, installs hook scripts, adds permission rules, creates skills and subagents, inserts marked blocks into `CLAUDE.md`, removes dead MCP servers — every write journaled with a backup.	`… apply` (`--param`, `--param-file recipe.name=path`, `--add FILE`)
Verify	Self-tests every installed hook and the status line, parses the settings files it writes (`--project` adds the project's `.claude/settings*.json`), checks CLAUDE.md markers. `~/.claude.json` and created skills/agents are checked by the `kaizen-verifier` agent.	`… verify`
Undo	Reverts a batch exactly, leaving anything you edited afterwards alone (`--force` reverts those too, after backing them up).	`… undo`

Everything runs on your machine; the one exception is the optional artifact console in the desktop app (see Privacy). The only thing a model ever sees is the digest (plus the short scrubbed prompt snippets stored beside it), and none of the agents reads your transcripts directly.

Install

As a plugin (recommended; adds /kaizen:* skills, three subagents and a weekly reminder):

/plugin marketplace add KROU4/claude-kaizen
/plugin install kaizen@claude-kaizen

Then, in any session:

/kaizen:start

As an agent kit (nothing to install — paste this into Claude Code):

Clone https://github.com/KROU4/claude-kaizen into a temporary folder, read its AGENTS.md, and run the Kaizen cycle for me.

From the shell (no Claude needed for the numbers):

git clone https://github.com/KROU4/claude-kaizen && cd claude-kaizen
node scripts/kaizen.mjs measure        # digest to stdout, files in ~/.claude/kaizen/ (`paths` prints the real location)
node scripts/kaizen.mjs recommend      # what the data says you should change
node scripts/kaizen.mjs console        # pick in the browser → choices.json
node scripts/kaizen.mjs choices --import my-choices.json   # validate a choices file you wrote or got back from an artifact
node scripts/kaizen.mjs apply          # journaled, reversible
node scripts/kaizen.mjs apply --recipe claude-md.corrections --param-file claude-md.corrections.rules=rules.txt   # multi-line value from a file
node scripts/kaizen.mjs verify         # self-test what was installed (--project DIR for project scope)
node scripts/kaizen.mjs undo --last

Or without cloning, straight from GitHub (pin a tag for repeatable runs):

npx github:KROU4/claude-kaizen#v0.1.12 measure
npx github:KROU4/claude-kaizen#v0.1.12 recommend

Run measure again a week later: the digest ends with a what moved table (interruptions, permission prompts, timeouts, compactions, corrections, cost…) comparing the two measurements.

Requirements: Node.js 18+ and Claude Code with some history (transcripts live for cleanupPeriodDays, 30 by default).

The console

Recommendations are grouped by category with their evidence, impact and risk. Tick what you want, adjust parameters (which formatter, deny or ask, which MCP servers), add a note for Claude ("tests are make check"), press Apply. In the Claude desktop app the page is a published private artifact backed by its database, so Claude reads your choices (after you reply "done" in chat) without you pasting anything, shows you the selection and asks for a confirmation before applying. In a plain terminal it is a page on 127.0.0.1 that writes choices.json and exits.

Recipes

The catalog lives in recipes/catalog.json; the generated reference is docs/recipes.md. Each recipe has a trigger over measured metrics, bilingual explanations, parameters, actions and an automatic undo. Highlights:

Permissions — a read-only allowlist generated from the commands you actually repeat (the /fewer-permission-prompts algorithm, with scope choice and interpreter blocklist); deny rules for .env, keys and SSH material; acceptEdits when you never reject edits.
Hooks — destructive-command guard, secret-file guard, format-on-edit (only when the formatter is installed), verify-before-stop (sends Claude back once when code changed without a test run), desktop/Telegram notifications, session-start git context, pre-compaction handoff notes, long-command nudge.
Rules (CLAUDE.md) — language, read-before-edit, autonomy, verification, and rules distilled from your own corrections, each in a marked block.
Skills and subagents — a skill from your most repeated prompt, a read-only code-reviewer, a Haiku test-runner that keeps logs out of your context.
Context budget — unused MCP servers (removed with backup), plugins none of whose skills you use, oversized CLAUDE.md, compaction pressure.
Advice — failing or slow hooks (with the culprit's name), model routing for subagents, plan mode, what /insights says breaks most.

Propose a recipe with the recipe issue template.

What gets measured

Sessions, prompts and tools per session · interactive permission prompts vs auto-accepted · rejections, interruptions, corrections · compactions and peak prompt size · tool errors by class (timeouts, edit mismatches, not-found, test failures…) · slow and repeated commands · hooks: runs, failures, latency per hook · subagent and skill usage · MCP servers configured vs used · installed vs invoked skills and plugins · CLAUDE.md load in tokens · tests per commit, destructive commands, force pushes · working hours, prompt length, repeated prompt openings · model mix and reported cost · /insights friction, outcome and satisfaction labels when available.

Full list with definitions: docs/recipes.md#metrics-recipes-can-test.

Privacy

Nothing leaves your machine, with two opt-in exceptions. There is no telemetry; the only network calls are the optional Telegram notification hook you configure yourself and the artifact console below.
Artifact console (desktop app only). The page is published as a private claude.ai artifact: the recommendations are part of the page, and your ticks, parameters and note are stored in the artifact's database until Claude reads them and deletes the document. Do not share the link and do not type secrets there (Claude asks for a Telegram bot token in chat instead). The local browser console and the terminal flow upload nothing.
The digest keeps counts and short, secret-scrubbed snippets of your own prompts (needed to spot repeated instructions). Tool output and Claude's replies are never retained.
State lives in ~/.claude/kaizen/ (override with KAIZEN_STATE_DIR; node scripts/kaizen.mjs paths prints the real location). It also holds the hook scripts and status line Kaizen installed and the undo journal, so do not delete it while anything is applied: run undo --all first, then delete the folder. Deleting it first leaves hooks in settings.json pointing at missing files (they fail on every event and stop guarding) and makes undo impossible.
Every change is journaled under ~/.claude/kaizen/applied/<batch>/ with backups of the files it touched.

More in the FAQ, architecture notes and the Claude Code formats reference.

How it compares

Kaizen was built after reading ~60 related projects (docs/prior-art.md). The analysis side is crowded (/insights, sniffly, ccusage, claude-code-doctor); the writing side is a set of single-surface tools (claude-reflect → CLAUDE.md, cc-safe-setup → hooks, /fewer-permission-prompts → allowlist, ccstatusline → status line). Kaizen is the loop: evidence → reviewed, reversible, multi-surface changes → measured again. Zero dependencies, Node only, Windows/macOS/Linux.

Repository layout

.claude-plugin/     plugin.json, marketplace.json (install from this repo)
skills/             /kaizen:start  measure  apply  undo  status  recipes
agents/             kaizen-analyst, kaizen-smith, kaizen-verifier
hooks/              plugin hook: weekly SessionStart reminder
scripts/            kaizen.mjs CLI + lib/ (scanner, metrics, recipes, apply, console)
recipes/            catalog.json + templates/ (hook scripts, rules, agents, status line)
console/            the interactive page (vanilla HTML, offline)
locales/            en.json, ru.json
docs/               recipes.md (generated), reference/, prior-art.md, assets/
tests/              node --test, synthetic fixtures only

Contributing

node scripts/kaizen.mjs selfcheck && npm test must pass. Recipes need en and ru text, a trigger, evidence and working undo. See CONTRIBUTING.md.

License

MIT