---
title: "How Claude finds agent skills it does not know: 93 runs, October 2026"
canonical_url: https://cookiesforai.app/research/how-claude-finds-agent-skills-october-2026
last_updated: 2026-10-01
type: article
summary: "In 93 runs on Claude Opus 5.5, Sonnet 5.5 and Fable 5.1 on 2026-10-01, every question about a little-known agent skill made Claude search, and 13 of 18 answers linked a skill directory page."
---

# How Claude finds agent skills it does not know

Updated 2026-10-01. First published 2026-10-01.

When Claude is asked for an agent skill it does not know, it searches, and the page that reaches the answer is usually one page per skill on a skill directory. In 93 runs on three Claude models on 2026-10-01, all 18 search runs on little-known skills searched the web, and 13 of the 18 answers linked a third-party skill directory in the answer itself. For famous skills, Claude went to the publisher's own repository instead.

## How we ran the test

| | |
|---|---|
| Date | 2026-10-01 |
| Models | claude-opus-5-5, claude-sonnet-5-5, claude-fable-5-1, through Claude Code 2.1.286 (API, web search on) |
| Setup | An empty folder, no user settings, no memory, only the web search and web fetch tools |
| Questions | 9 plain questions, no instructions: 3 about AI tool directories, 3 about little-known agent skills, Instagram posting, new coding tools, MCP servers for web research |
| Runs | 93: 2 to 3 search runs and 1 memory-only run per model and question, plus extra repeats for stability |
| What we counted | Links in the answer body. Claude Code's search tool also appends a Sources list to every answer, which we did not count as a citation |

The three skill questions were: "Is there an agent skill that helps my coding agent review Terraform plans?", "What agent skills exist for making videos with code?" and "Find me an agent skill for writing database migrations safely."

## Results

**Little-known skills go to skill directories.** All 18 search runs searched (Sonnet 5.5 too, which skipped search on most other questions). The answer bodies linked: skills.lc 6 times, mcpmarket.com 4, explainx.ai 3, agent-skills.md 3, agentskill.sh 2, lobehub.com 2, crossaitools.com 2, skillstore.io 2, and seven others once. GitHub repositories of the skills themselves were linked in 17 of 18 runs. The pages that won had one page per skill with the skill's name in the address.

**Claude searched in the user's words plus "SKILL.md".** Typical query: "agent skill Terraform plan review SKILL.md", often with "Claude Code" added.

**Broad directory questions were decided from memory.** There's An AI For That was named in 27 of 27 search runs and 9 of 9 memory-only runs of the three directory questions, and Futurepedia in all 27 search runs.

**A precise attribute let new sites in.** "Which AI tools directory has the widest range of categories?" made all 9 runs search, and two sites no memory run named entered the answers: tooldirectory.ai in 8 of 9, with a number in its title, and aicentralresources.com in 7 of 9.

**Pages were rarely opened.** Across this test and an earlier one (147 runs), Claude opened a page in 55 of 489 cases where a domain was cited; most answers were written from search result titles and memory. Titles shaped as "best" or "top" lists reached the answer body 4% of the time, against 20% for other titles.

**The head is stable; the order is not.** Over 15 runs on MCP servers for web research, Brave Search, Exa, Tavily, Firecrawl, Fetch and Playwright were named every time, but the first pick split Tavily 6, Brave 4, Exa 3, Fetch 2.

**YouTube barely appears.** 2 of 1,483 search results Claude was shown were YouTube videos, and none reached an answer.

## What this means

For an agent skill that a model does not already know, the way into an answer is a page about that one skill, named with the skill's name and the task words people use, on a site the search engine already shows. This is why CookiesForAI has one page per skill with its name in the address.

## Limits

Claude only, through Claude Code and the API; the claude.ai app may cite differently. Small samples: 2 or 3 search runs per model and question. Results hold for these three models on this date; newer models may behave differently. The same questions on ChatGPT, Gemini, Perplexity, Google and Grok are in [our six-engine study](/research/ai-assistants-field-study-october-2026).