claude

self-improvement

Scores every Claude Code session for the lessons it holds and proposes where each one belongs in the harness: a watermarked background scan kicked by the SessionStart hook, one haiku call per dirty session, then a suggestion pass over the existing skill inventory. Browsable in the self-improve TUI. High scorers are archived for the capture-lesson skill to turn into skills and the dream skill to consolidate.

2 on GitHubby popoffvgv0.38.0Apache-2.0updated 5 days ago

Not in a marketplace yet, so it loads for one session with --plugin-dir.

Plugin self-improvement in popoffvg/dotfiles. Commands read from the repo's manifest — 4 days ago.

This plugin runs code on your machine — its hooks fire automatically on events like tool use. The exact commands are listed below; read them before installing.

What it installs

Skills (3)

  • capture-lesson

    Route a captured lesson to the cheapest container that enforces it — a hook, a lefthook job, a CLAUDE.md line, engram, or a skill when the lesson is about how a request is read. Use after a user correction, after the user states how they want a task done or how work is done in this repo, or after yo

  • dream

    Consolidate the accumulated skill corpus — global (~/.claude/skills) and project (.claude/skills). Run periodically to prune dead rules, unite overlapping ones, and generalize the over-specific. The complement to capture-lesson: that skill captures, this one consolidates.

  • lessons

    Report the session archive the self-improvement scan has collected — what the backlog is made of, which clusters it falls into, and what /dream would work on next. Use when the user asks what lessons the scan stored, what the statusline lesson count is made of, or which lessons are worth harvesting.

Hooks (2)

  • PreToolUse

    ${CLAUDE_PLUGIN_ROOT}/hooks/allow-self-improve-writes.sh

  • SessionStart

    ${CLAUDE_PLUGIN_ROOT}/hooks/session-start.sh

View source on GitHub

More from popoffvg/dotfiles

All 8 plugins →

wm

claudecodex

Work lifecycle management v2 — phased workflows (research, plan, implement) with agents, commands, and hooks.

18 skills · 34 commands · 20 agents · hooks · LSP

2

harness-dev

claude

Authoring aids for the Claude Code harness — skill shapes (workflow, loop, instruction, router), golden words that stand in for a rule set, pruning an existing skill or doc corpus into the house style, repo bootstrap (README, CLAUDE.md, deploy and adr skills, settings), and agent-work metrics from a per-turn ledger (re-ask ratio and causes, prompts to done, pasted blocks that should be a skill).

8 skills · 2 agents · hooks

2

line-comment

claude

Review comments on lines of any file, kept outside the file. Zed writes them through the line-comment language server; /line-comment:act reads the store and does the work, /line-comment:write places comments, /line-comment:unresolved reports what is still open, /line-comment:prune throws comments away, /line-comment:distill turns the habits act saved into review rules.

5 commands

2

hi

claude

Per-session prompt stack: push/pop/list/drop prompts across a session.

4 commands

2

smart-commit

claude

Split the working tree of every nearby git repo into logical commits at hunk granularity. A sonnet subagent proposes the split into an editable file, the operator answers each block, a second subagent builds the commits in a throwaway worktree and cherry-picks them back.

1 skill · 1 command · 1 agent

2

prompt-stack

claude

Push, pop and list prompt drafts in files under ~/.claude/prompt-stack, from a key-opened menu

2

Maintain self-improvement?

This page updates itself from your repo. Add the badge to your README, or get it Featured to pin it above the organic results with a followed link to your site.

claude & codex plugins8 plugins