Small, sharp, agent-agnostic skills that keep an AI agent's work honest.
A whetstone doesn't make a tool — it keeps its edge true. Whetstone is a growing collection of tiny open-source skills for AI agents: not frameworks, but disciplines that keep an agent's work on-track, high-quality, and easy to hand off.
- Small and self-contained. One folder, one job. No build step, no install, no runtime.
- Agent-agnostic. Works as a plain prompt in any agent — Claude, OpenAI, Gemini, Cursor, a local model — and as a skill / slash command where those exist. Nothing assumes a particular platform; every capability degrades gracefully.
- Prompt-first, optional tooling. The discipline lives in plain language. Any helper script is optional, zero-dependency, and never required.
- Honest by construction. Checks are anchored to structure and substance, not vibes — a green light has to mean something.
- Plain words over jargon. If a reader has to learn a new token for an idea they already know, that's a defect, not concision.
| Skill | What it does | Works with any AI / model |
|---|---|---|
| Throughline | Task-Track & Handoff for AI agents on long, multi-session work — within a session, one ⭐️ MAIN goal anchor in an emoji-tagged task list so reasoning doesn't drift; across sessions or agents, a plain-markdown handoff the next agent re-derives before trusting. | ✅ Yes — prompt-first, agent-agnostic |
| Session Measurement | Score an agent session by session and read the trend across versions — so a model / frame / skill change shows up as a real regression or improvement, not a vibe. Six honest counts per session, plain-markdown trend table (optional charts), any agent. | ✅ Yes — any agent |
| notify-me | Calls you back when an agent needs you — the moment it gets blocked waiting on your decision, it pushes to your phone (via ntfy) and/or chimes on your computer, so you can step away and still know when you're needed. Hook-based, no account, no background process. | ✅ Yes — any agent with hooks |
| more coming | — | — |
Whetstone is a Claude plugin marketplace, so you can install these skills directly in Claude Code or Cowork:
/plugin marketplace add jovesun-lab/whetstone
/plugin install session-measurement@whetstone
/plugin install throughline@whetstone
/plugin install notify-me@whetstone
Installed skills are enabled by default and trigger automatically when relevant. Pull future updates
with /plugin marketplace update whetstone. Once Whetstone is accepted into Anthropic's plugin
directory, you'll also be able to find and install these from the in-app Plugins directory
(Customize ▸ Plugins) without adding the marketplace by hand.
Each skill is a self-contained folder. Drop it where your tool reads skills / commands
(Claude Code, Cursor, …), or just paste the relevant file into any agent as a prompt. Start with
each skill's own README.md.
Each Whetstone skill is built to be cheap to keep installed and modest to use (per skill, approximate, tokenizer-dependent):
- Installed, idle: ~250 tokens — only the trigger description sits in context.
- When it fires: a few thousand tokens — the skill body plus whatever it produces (e.g. a Throughline session that writes one handoff runs ~3–5k).
- Optional scripts / validators: 0 model tokens — they run as code.
On progressive-disclosure hosts (Claude Code, Cursor) you pay the ~250 baseline and spend the rest only when a skill actually fires — and triggers are scoped so short, one-pass tasks don't load them.
Every Whetstone skill is MIT unless noted otherwise. Free to use, modify, embed, redistribute.
Whetstone is by guxcom. Feedback, issues, and PRs welcome on any skill.