claude-help

name	claude-help
description	Search and read Claude Code official docs, changelog, and GitHub issues. Use when the user asks how Claude Code works, what a setting/env var/flag does, wants release notes, or wants to find/triage a known issue. Handles proxy + GitHub fallback automatically for network problems.
user-invocable	true

Claude Help

Search and read authoritative Claude Code information: official documentation, the changelog, and GitHub issues in anthropics/claude-code. Answers "how do I…", "what does X do", "is this a known issue", and "what changed in version Y" questions.

When to Use

• User asks how Claude Code works (features, hooks, slash commands, settings, env vars, flags, MCP, IDE integrations).
• User asks what a specific setting / env var / CLI flag means.
• User wants release notes or what changed in a version.
• User hit an error / odd behavior and wants to find a known issue.
• User asks to search Claude Code GitHub issues (bugs, feature requests, questions).

CRITICAL: Network, Proxy & Fetcher Rules

There are three independent fetchers — they fail differently. Know which one you're using.

Fetcher	Tool	Needs proxy?	Notes
Shell network	Bash (gh, curl, git)	Yes for github.com / pypi / npm	Env vars do NOT persist between Bash calls. Do NOT hardcode proxy commands here — read the /proxy skill for the correct incantation.
Built-in fetch	WebFetch / WebSearch	No	Has its own network path; proxy env vars do nothing for it. Geo-blocked on some platform.claude.com URLs.
Exa MCP	mcp__exa__web_search_exa / mcp__exa__web_search_advanced_exa / mcp__exa__crawling_exa / mcp__exa__get_code_context_exa	No	Fetches via Exa's own infra. Bypasses the platform.claude.com geo-block (verified) and gives cleaner search with highlights/summaries. Preferred over WebSearch when available.

So:

• Reading docs via WebFetch or Exa crawl → just call it. No proxy.
• Searching issues / changelog via gh or curl → proxy is required. Read the /proxy skill for the correct command form (when proxy is needed, how to enable it, and the reminder that each Bash call is a fresh shell). Do not inline proxy on here — the /proxy skill is the source of truth.
• If a Bash network command fails, the most common cause is proxy not enabled or a flaky proxy. Consult the /proxy skill and retry once with proxy explicitly enabled before assuming the endpoint is down.

Known geo-block + the Exa bypass (important)

code.claude.com/docs/en/* is reachable via WebFetch. But platform.claude.com/docs/* (where the model-overview / max-output-tokens table lives) 307-redirects to "app unavailable in region" under WebFetch in some networks.

Exa's crawler bypasses this — mcp__exa__crawling_exa fetches platform.claude.com pages in full (verified). So when WebFetch reports that 307 redirect, do not chase it — re-fetch the same URL with mcp__exa__crawling_exa instead, then fall back to gh/Exa search only if Exa also fails.

Which Exa tool to use

Tool	Use for
mcp__exa__crawling_exa	Reading a known URL in full (especially a geo-blocked one). Batch up to several URLs per call.
mcp__exa__web_search_exa	Natural-language search — describe the ideal page, not keywords. Best default search.
mcp__exa__web_search_advanced_exa	Search with filters: domains, date range, labels. Use includeDomains: ["anthropic.com","code.claude.com","github.com"] to restrict to official sources. ⚠️ Can return huge payloads — pass numResults ≤ 5 and textMaxCharacters to stay manageable.
mcp__exa__get_code_context_exa	API/library usage examples & docs — use when the question is about the Claude API/SDK specifically.

Exa tools are only present if the Exa MCP server is connected. If the tools aren't available, silently fall back to WebSearch/WebFetch.

Sources, in order of preference

Try the cheapest reliable source first; escalate only if it's empty.

1. Local changelog cache — ~/.claude/cache/changelog.md (~418k, offline, always works). Best for "what changed in version X" / "when was feature Y added". No proxy needed.
2. Official docs via WebFetch — https://code.claude.com/docs/en/<page>.md (note the .md suffix for clean markdown). Best for how-to / settings / env vars / flags. No proxy needed. If WebFetch geo-blocks (307 on platform.claude.com), re-fetch with mcp__exa__crawling_exa.
3. GitHub issues via gh — anthropics/claude-code. Best for known bugs / behavior questions / confirming whether something is a reported issue. Needs proxy (see /proxy).
4. Web search via Exa (mcp__exa__web_search_exa, preferred) or WebSearch (fallback) — for recent info, when docs/gh are blocked, or to find a known URL before crawling it. No proxy needed.

Workflows

A. "What does X do?" (setting / env var / flag / feature)

1. Check the local changelog first if it sounds version-recent:

   grep -n -i "<term>" ~/.claude/cache/changelog.md | head -20
   

2. Fetch the relevant docs page (no proxy):

   WebFetch url="https://code.claude.com/docs/en/settings.md" prompt="What does <X> do?"
   

   Common pages: settings, env-vars, cli-reference, hooks, slash-commands, mcp, model-config, costs, statusline, ide-integrations.
3. If WebFetch geo-blocks the URL (307 redirect on platform.claude.com), re-fetch the exact same URL with mcp__exa__crawling_exa — it bypasses the block (verified).
4. Delegate for hard "how do I / does Claude" questions: spawn the claude-code-guide agent (purpose-built for Claude Code / SDK / API Q&A). It can search docs + release notes for you.
5. If docs are geo-blocked or empty after the Exa retry, fall back to Exa/WebSearch, then to grepping the changelog.

B. "What changed in version X?" / release notes

1. Grep the local changelog (offline, fast):

   grep -n -i "^## 2.1.<X>\|<feature-term>" ~/.claude/cache/changelog.md | head -30
   

2. Present the matching ## <version> entry. Quote it; don't paraphrase behavior.

C. "Is this a known issue?" / search GitHub issues

This is the workflow that needs proxy + gh together — the network problem the user mentioned.

Do NOT hand-roll the gh commands or proxy incantation here. Read the /github skill — it is the source of truth for:

• proxy setup + the connectivity check sequence (it composes /proxy internally),
• the exact gh issue list / gh issue view / gh api search/issues commands and their qualifiers,
• the present-findings and create-issue (approval-gated) steps.

Workflow using those skills:

1. Read /github to load the issue-search workflow (and let it handle proxy via /proxy).
2. Target repo: anthropics/claude-code (unless the user names another).
3. Extract keywords from the user's description: error text, flag/function names, concepts. Drop stop words.
4. Search open + closed issues using the commands from /github (closed issues often carry the fix). Use --repo anthropics/claude-code.
5. Read a specific issue + comments using the read-issue step from /github, scoped to --repo anthropics/claude-code and --comments for the full discussion.
6. Present findings: exact matches (link + state), related issues, any known workaround from closed issues, and a verdict (known? open/closed? fix merged?).
7. If nothing matches: try broader keywords, then say it may be unreported and offer to draft a new issue — never create/comment on an issue without explicit user approval (per /github).

D. Network failure recovery

If a gh command fails:

1. Was proxy enabled for that Bash call? Each Bash call is a fresh shell — consult /proxy for the correct enable command and re-run with proxy on.
2. Still failing? Use /proxy's status check to diagnose, and enable explicitly — avoid the toggle alias (it can flip proxy off by accident). See /proxy for the exact commands.
3. If proxy is up but GitHub still fails, search via Exa (mcp__exa__web_search_exa) for the issue with the error text + "claude code issue" (no proxy needed); fall back to WebSearch if Exa is unavailable.
4. For geo-blocked doc URLs, re-fetch with mcp__exa__crawling_exa (verified bypass) before giving up.
5. As a last resort, suggest the user open the docs/issue URL directly in a browser.

Answering

• Cite the source (docs URL, changelog version line, or issue number + link) for every non-obvious claim.
• If a source was unreachable (geo-block / network), say so explicitly and note which source you used instead — don't silently substitute.
• Quote docs/changelog verbatim for exact behavior; paraphrase only for summary.
• If the answer is model-dependent or version-dependent and you couldn't pin down the exact number, say that plainly rather than guessing.

Notes

• This skill orchestrates; it does not duplicate commands. For proxy enablement and the exact gh/issue workflows, read the /proxy and /github skills — they are the source of truth and stay current. Inlining their commands here would drift.
• Three fetchers, three rules: WebFetch/WebSearch and Exa MCP = no proxy; gh/curl/git in Bash = proxy required (see /proxy), every call.
• Exa MCP bypasses the platform.claude.com geo-block that WebFetch hits — use mcp__exa__crawling_exa to re-fetch any URL WebFetch 307-redirects. If Exa tools aren't connected, silently fall back to WebSearch.
• Local changelog at ~/.claude/cache/changelog.md is the fastest, most reliable source for version-specific questions — use it first.
• code.claude.com/docs/en/* works via WebFetch; platform.claude.com/docs/* is geo-blocked under WebFetch but reachable via Exa.
• Exa web_search_advanced_exa can return very large payloads — always pass numResults ≤ 5 and textMaxCharacters, and prefer includeDomains to official sources.
• Delegate deep Claude Code / SDK / API Q&A to the claude-code-guide agent when a single docs page won't suffice.
• Never create or comment on anthropics/claude-code issues without explicit user approval.
• Composes: /proxy (proxy enablement) and /github (gh + issue workflows) for the underlying commands.