What is danshari.skill?

An agent skill whose job is to audit the other skills you have installed — finding the ones your current model already knows how to do, your harness already ships as a built-in, or your MCP servers already cover, and archiving them so you stop paying their per-turn description tokens.

Why delete skills at all? What is the harm in keeping them?

Every installed skill puts its name and description into the system prompt on every single turn — a standing cost. Descriptions also compete for triggers, and a wrong trigger is worse than no skill at all. And skills rot: hard-coded paths, APIs and model names expire, after which each trigger actively misleads.

Will it delete something I actually need?

Nothing is deleted. You review the verdict table line by line and can veto any of it. Knowledge skills — private facts, accounts, internal conventions — are a protected red line. Every removal is a move into an archive directory with a MANIFEST, restorable with one command. rm is never executed.

When should I run an audit?

After a model upgrade, after a major harness release, after connecting new MCP servers, or 90 days since the last audit. Model release day is skill decluttering day.

Which agents does it support?

Any host that reads SKILL.md — Claude Code, OpenClaw, Codex, Hermit. If your skills live somewhere non-standard, point tools/inventory.py --root at the directory.

断舍离.skill · MIT · zero dependencies

The first skill that tells you to delete skills.

The whole ecosystem is addition — awesome lists collect thousands of skills, single repos ship three hundred. Nobody does subtraction. But skills are not free: every one you install bills you tokens on every turn, and each time the model gets a generation stronger, another batch of them turns from asset into liability.

danshari.skill cuts what teaches the model to do what it already knows,
and keeps what tells the model what it could not possibly know.

View on GitHub → See a real audit

git clone https://github.com/swaylq/danshari-skill.git ~/.claude/skills/danshari-skill

Demo audit: 12 → 5
Description tax: −66%
Dependencies: 0
rm calls made: 0

Why subtract

Skills are not free — you pay three costs

There is always a reason to install one, and never an occasion to remove one. So the directory only grows, while the model keeps getting stronger.

Token tax

Its name and description enter the system prompt on every turn. A dozen skills is two thousand tokens per turn, billed whether or not today's task needs any of them.

Mis-triggering

Descriptions compete for the same work. One wrong trigger is worse than having no skill at all — the model marches off confidently holding an obsolete manual.

Rot

Hard-coded paths, APIs and model names expire. A rotten skill misleads every time it fires, and it will never tell you that it has gone stale.

What it looks like

Not "delete a few that look untidy"

Every skill is tried against the model, harness and MCP servers you are running right now, and judged on evidence.

~/.claude/skills — danshari.skill

you       ❯ new model just landed — declutter my skills

danshari  ❯ Environment: claude-fable-5 · Claude Code (built-in search /
            subagents / auto-memory / plan mode) · MCP ×3 (slack /
            github / playwright)

            Inventory: 12 skills, ~2,160 tokens/turn of description tax.

            Blind test on think-harder: a subagent without it debugged a
            real race condition at equal quality → the model has
            internalised this. 210 tokens/turn paid for nothing.

            Verdicts: ✂️ remove ×6 · 🔧 trim ×1 · 🔗 merge ×1 ·
            ⏳ probation ×1 · ✅ keep ×2
            Evidence per line in the report — execute?

you       ❯ go ahead, but keep the selector script inside browser-scrape

danshari  ❯ Understood: script extracted to scripts/ first, shell archived after.
            Archived 7 (restorable), trimmed 1 (58 lines → 9), merged 1 pair.

            12 → 5 skills, ~1,425 tokens/turn saved (−66%).
            Restore: tools/archive.sh --restore <name>

Full report with the evidence behind every verdict: examples/2026-08-17-demo/report.md

How it judges

Triage by type, prove with five questions, confirm by blind test

"Stronger models make skills useless" is only true of the first type. A cleanup that cannot tell the types apart deletes exactly the wrong ones.

Type	Definition	Fate
Capability	General method — teaches the model how to think or work	Model upgrades obsolete it — the audit's focus
Environment	Wraps a tool, or fills a gap in the harness	Harness and MCP evolution replace it
Knowledge	Private facts: accounts, paths, internal conventions	Never obsoleted by a stronger model — protected

Q1Model already internalised it?

Q2Harness ships it built-in?

Q3An MCP server took it over?

Q4Absorbed by / duplicating another?

Q5Has it rotted?

Five verdicts, not two

Before any capability skill is removed, a subagent that does not load it attempts the skill's most typical task. Only equal quality justifies the cut. Evidence, not intuition.

Verdict	Meaning
✂️ Remove	Evidence is conclusive → archived (restorable). Never `rm`.
🔧 Trim	Mixed type: drop the obsolete lecture, keep the valuable facts — 200 lines often become 10.
🔗 Merge	Duplicates fighting over the same triggers, folded into the better one.
⏳ Probation	Evidence insufficient — not decided by force, flagged for the next audit.
✅ Keep	It earns back its own token tax.

Safety

A tool that touches your skills directory is only as good as its safety design

✓Never rm — every removal is a move into an archive directory. --restore brings it back in one command; the MANIFEST keeps the trail.
✓Nothing without confirmation — you see the verdict table first, and can veto line by line.
✓Knowledge skills are a red line — private facts, team policy, referenced scripts: kept no matter how strong the evidence.
✓Frisked before archiving — obsolete skills often bury facts that are not obsolete (URLs, accounts, scripts). Those are extracted first.
✓Every verdict carries evidence — the blind-test record, the named replacement (which built-in, which MCP server), proof of rot. Assertions don't count.
✓Self-judgment clause — danshari itself costs you ~180 tokens/turn and has to earn that back. If your model audits just as well without it, declutter danshari too.

Install

One clone, any host

Once installed, just say "declutter my skills" or "new model, clean up the skills" in your agent.

git clone https://github.com/swaylq/danshari-skill.git ~/.claude/skills/danshari-skill

Host	Target path
Claude Code	`~/.claude/skills/danshari-skill`
OpenClaw	`~/.openclaw/skills/danshari-skill`
Codex	`~/.codex/skills/danshari-skill`
Hermit	`~/.hermit/skills/danshari-skill`

Only after the methodology? Feed SKILL.md and references/rubric.md straight to your agent.