claude-skills

claude-skills

Skills for Claude Code — small, self-contained tools that teach Claude repeatable workflows.

🌳 dialog-tree

You're working through a hard problem with Claude. You ask a question — the answer raises three follow-ups. You dive into the first one; it branches again. Twenty minutes deeper, you surface and try to remember what else you meant to explore. It's gone. And once the session's context gets compacted, it's gone for Claude too.

dialog-tree keeps that map for you — as a live, visual tree you open in your browser: every branch of the conversation clickable, with what's answered, what's still open, and what you decided to skip. Claude maintains it incrementally; a zero-dependency HTML viewer in your repo renders it; your resolve/delete marks persist in the browser. It survives context compaction — and your own memory limits.

The tree is logical, not literal: nodes are topics, steps, and decisions distilled from the conversation — not raw chat messages. That's what makes it a map you can navigate, and what sets it apart from tools that visualize the message log itself.

📖 Full docs — data model, naming, localization, how updates work.

Install

Via plugin marketplace (recommended):

/plugin marketplace add ikotelkin/claude-skills
/plugin install dialog-tree@ikotelkin-skills

Or manually:

git clone https://github.com/ikotelkin/claude-skills.git
cp -r claude-skills/skills/dialog-tree ~/.claude/skills/

Then, in any Claude Code session: "Set up a dialog tree for this conversation" — Claude builds the tree and hands you a localhost link to open (in Claude Code it appears right in the built-in browser pane).

🧠 working-memory

A long session with Claude runs out of context and gets compacted — the transcript is squeezed into a summary, and the details go with it: what you decided, the edge-cases, the approaches you already ruled out. Claude keeps working on the thinned-out version, quality slips, and the next session opens blind — didn't we already do this?

working-memory gives each project a WORKING.md in the repo — a small, living file holding the current stage's state (decision log, dead ends, next steps). It survives compaction and new sessions; at stage end its settled conclusions are distilled up into long-term memory. Three levels, one-way flow — context → WORKING.md → long-term memory — kept apart by a single question: "when this stage is reset, should this line survive?"

📖 Full docs — the three levels, the belonging test, the checkpoint/consolidate cycle.

Install

Via plugin marketplace (recommended):

/plugin marketplace add ikotelkin/claude-skills
/plugin install working-memory@ikotelkin-skills

Or manually:

git clone https://github.com/ikotelkin/claude-skills.git
cp -r claude-skills/skills/working-memory ~/.claude/skills/

Then, in a Claude Code session: "Set up working memory for this project." After that, "checkpoint" saves state, and "consolidate — this stage is done" promotes the keepers and clears the slate.

💸 frugal-subagents

You're on Claude Fable 5.1 — the strongest model, with the tightest usage limit — and you ask Claude to find you an apartment. It splits the job between several helpers — one per listings site — and each runs on the same expensive model you're paying for. One helper decides its slice is big and starts five helpers of its own. Ten minutes later your session limit is gone, most helpers stopped mid-way, and what they found went with them. You tell Claude "use the cheaper model for searching, no helpers starting helpers" — and it holds until the conversation gets long, its memory is compressed, and the instruction quietly disappears.

frugal-subagents builds the instruction into Claude Code itself, so it can't be forgotten: helpers run on a cheaper model unless Claude specifically asks for a stronger one (and Claude is told which model each helper got), helpers can't start helpers, and there are at most 12 per session — when a limit kicks in, Claude says why and works with what it has. Two ready-made cheap helpers come along: web-scout for searching and reading many pages, extractor for turning the findings into a clean table or list; both save their work to files and hand Claude a short summary.

📖 Full docs — what changes after install, how to tune the defaults, how it works under the hood.

Install

/plugin marketplace add ikotelkin/claude-skills
/plugin install frugal-subagents@ikotelkin-skills

Start a new session — no settings to change. The guard needs Node.js; if your computer doesn't have it, Claude will notice and offer to install it for you.

Repo layout

skills/
  <skill-name>/
    SKILL.md        ← instructions Claude follows
    README.md       ← human-facing docs for this skill
    assets/         ← templates the skill copies into your project
    agents/         ← subagent definitions the plugin ships (if any)
    hooks/          ← hooks the plugin ships (if any)

Each skill is independent — take only what you need.

License

MIT