Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →What this skill does and when to use it. Claude reads this to decide relevance. Include keywords users would naturally say.
.claude/skills/lukerenton-my-skill/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 13% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 44% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 41% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 1% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 17% | 0% |
These two fields are the only ones required. name must be lowercase with hyphens, max 64 characters, and match the parent directory name. description is what Claude reads at startup to decide when the skill is relevant (max 1024 characters).
Add any of these to the --- block above to customise behaviour:
| Field | Example | Purpose | |---|---|---| | argument-hint | [issue-number] | Hint shown during autocomplete to indicate expected arguments | | disable-model-invocation | true | Prevent Claude from auto-loading. User must type /name explicitly. Use for deploys, sends, destructive ops | | user-invocable | false | Hide from the / menu. Claude can still load it automatically. Use for background knowledge | | allowed-tools | Read, Grep, Bash(npm *) | Tools Claude can use without asking permission. Space-delimited, supports patterns | | model | claude-sonnet-4-6 | Override the model when this skill is active. Useful for cost control | | context | fork | Run in an isolated subagent. Skill content becomes the subagent's prompt | | agent | Explore | Which subagent runs when context: fork. Built-in: Explore, Plan, general-purpose, or custom from .claude/agents/ | | license | Apache-2.0 | License name or reference to a bundled LICENSE file | | compatibility | Requires git, docker | Environment requirements (max 500 chars) | | metadata | key-value pairs | Arbitrary metadata (author, version, etc.) |
Everything below the frontmatter is the instruction body. Claude reads this when the skill is activated. Write whatever helps Claude perform the task. There are no format restrictions.
Good body content includes:
Placeholders are replaced with real values before Claude sees the content:
| Placeholder | Resolves To | |---|---| | $ARGUMENTS | Everything the user typed after the skill name | | $ARGUMENTS[N] or $N | A specific argument by index (0-based) | | ${CLAUDE_SESSION_ID} | The current session ID | | ${CLAUDE_SKILL_DIR} | Path to this skill's directory |
Example: /my-skill SearchBar React Vue gives $0 = "SearchBar", $1 = "React", $2 = "Vue".
If $ARGUMENTS is not present in the content, arguments are appended as ARGUMENTS: <value>.
The ! backtick syntax runs shell commands before the content reaches Claude. Output replaces the placeholder:
gh pr diff cat package.json | jq .dependencies git diff --name-only This is preprocessing. Claude only sees the final output, not the commands.
Keep SKILL.md under 500 lines. Move detailed material to separate files and reference them from the body:
Use relative paths from SKILL.md. Keep references one level deep. Navigate into these folders to learn more about each.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 8,652 | 11,059 | +28% | 1 | 1 | 0% | 1,487 | 1,680 | +13% | 0 | 0 | — |
case-02 | fail→pass | 9,731 | 7,019 | -28% | 1 | 1 | 0% | 1,447 | 2,089 | +44% | 0 | 0 | — |
case-03 | fail→pass | 15,113 | 15,040 | -0% | 1 | 1 | 0% | 2,370 | 3,334 | +41% | 0 | 0 | — |
case-04 | pass→pass | 6,358 | 2,777 | -56% | 1 | 1 | 0% | 954 | 1,328 | +39% | 0 | 0 | — |
case-05 | pass→pass | 10,124 | 3,744 | -63% | 1 | 1 | 0% | 1,584 | 1,510 | -5% | 0 | 0 | — |
case-06 | fail→fail | 13,050 | 2,192 | -83% | 1 | 1 | 0% | 2,095 | 1,279 | -39% | 0 | 0 | — |
case-07 | fail→pass | 7,479 | 1,978 | -74% | 1 | 1 | 0% | 1,180 | 1,190 | +1% | 0 | 0 | — |
case-08 | fail→pass | 8,197 | 3,093 | -62% | 1 | 1 | 0% | 1,195 | 1,403 | +17% | 0 | 0 | — |
case-09 | fail→pass | 13,819 | 1,777 | -87% | 1 | 1 | 0% | 659 | 1,162 | +76% | 0 | 0 | — |
case-10 | pass→pass | 13,799 | 1,724 | -88% | 1 | 1 | 0% | 2,125 | 1,171 | -45% | 0 | 0 | — |
case-11 | pass→pass | 5,897 | 2,464 | -58% | 1 | 1 | 0% | 864 | 1,284 | +49% | 0 | 0 | — |
case-12 | pass→pass | 9,997 | 2,550 | -74% | 1 | 1 | 0% | 1,512 | 1,294 | -14% | 0 | 0 | — |
case-13 | fail→pass | 19,436 | 1,841 | -91% | 1 | 1 | 0% | 1,701 | 1,183 | -30% | 0 | 0 | — |
case-14 | fail→pass | 8,907 | 2,127 | -76% | 1 | 1 | 0% | 1,279 | 1,204 | -6% | 0 | 0 | — |
case-15 | fail→pass | 12,639 | 2,792 | -78% | 1 | 1 | 0% | 1,747 | 1,327 | -24% | 0 | 0 | — |
case-16 | fail→pass | 14,078 | 2,891 | -79% | 1 | 1 | 0% | 419 | 1,251 | +199% | 0 | 0 | — |
case-17 | fail→pass | 8,022 | 2,965 | -63% | 1 | 1 | 0% | 765 | 1,321 | +73% | 0 | 0 | — |
case-18 | pass→pass | 11,112 | 2,335 | -79% | 1 | 1 | 0% | 1,624 | 1,265 | -22% | 0 | 0 | — |
case-19 | fail→pass | 11,369 | 2,315 | -80% | 1 | 1 | 0% | 1,656 | 1,260 | -24% | 0 | 0 | — |
case-20 | pass→pass | 17,550 | 18,453 | +5% | 1 | 1 | 0% | 3,043 | 4,130 | +36% | 0 | 0 | — |
case-21 | pass→pass | 12,499 | 12,406 | -1% | 1 | 1 | 0% | 2,128 | 2,943 | +38% | 0 | 0 | — |
case-22 | pass→pass | 13,833 | 13,408 | -3% | 1 | 1 | 0% | 2,197 | 3,183 | +45% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 20 counted toward the lift figure. The other 2 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +55 percentage points is the difference between those two pass rates over the 20 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.