ask-claude-code-doc-agent

---
name: ask-claude-code-doc-agent
description: "Ask questions about Claude Code features, commands, workflows, and best practices. Dispatches the built-in claude-code-guide agent to search official docs. Use when: the user asks how a Claude Code feature works, wants to know which feature solves a problem, or has a real-world workflow problem that Claude Code features might solve. Handles three levels: direct lookups, problem-to-feature mapping, and complex problem-to-solution synthesis."
---

# Ask Claude Code Doc Agent

Dispatch the built-in `claude-code-guide` agent to answer questions about Claude Code.

## Detect the Level

Before dispatching, classify the question:

**Level 1 — Direct lookup:** User asks how a specific feature works.
Examples: "How does session forking work?", "What does --continue do?", "Explain hooks."

**Level 2 — Problem-to-feature mapping:** User describes a want/need that maps to a feature.
Examples: "I want to work from my phone", "How do I share rules across my team?", "Can Claude run in CI?"

**Level 3 — Problem-to-solution synthesis:** User describes a messy real-world problem. No single doc page answers it. The answer requires composing multiple features together.
Examples: "I have 4 projects and switching between them is overwhelming", "I want my teammate to pick up where I left off on a task."

## Level 1 and 2: Dispatch Directly

Dispatch to the `claude-code-guide` agent with the user's question. Include this instruction in the prompt:

> Return only what directly answers the question. No preamble, no "here's what I found." Maximum 10 lines unless the answer genuinely requires more. Prefer a command example over a paragraph of explanation. If there are multiple approaches, state the simplest one first.

After the agent returns, distill its response further if needed. The user wants:
- The command or flag, not the essay
- One clear recommendation, not a menu of options
- Real examples, not abstract descriptions

## Level 3: Interview First

Do NOT dispatch the agent immediately. The user's problem is ambiguous and the agent will overengineer the answer without clarity on intent.

Use AskUserQuestion to ask ONE focused clarifying question. Frame it as narrowing down the constraint that matters most. Examples:
- "Are these projects in the same repo or separate repos?"
- "Do you need them running simultaneously or one at a time?"
- "Is the goal to reduce setup effort or reduce ongoing management effort?"

After the user answers, dispatch to `claude-code-guide` with both the original problem AND the clarified constraint. Include the same conciseness instruction from Level 1/2.

After the agent returns, synthesize — don't just relay. Connect the features into a workflow. State it in under 10 lines.

If the synthesized answer still feels too complex, it probably is. Ask yourself: is there a simpler combination of features that solves this? The best Claude Code solutions use 1-2 features, not 5.

## Output Format

Always end with the actual commands the user would run. No trailing summaries, no "let me know if you have questions." Just the answer and the commands.