FAQ & Troubleshooting
“Claude keeps forgetting something I told it earlier in the session.”
Section titled ““Claude keeps forgetting something I told it earlier in the session.””Likely the context window filled up and got compacted, or you’ve been running a long session that should’ve been split into smaller ones. See Sessions, Context & Memory — start a fresh session per feature, and put anything that must always be true into CLAUDE.md rather than relying on it staying in a long conversation’s working memory.
“A skill I wrote isn’t triggering when I expect it to.”
Section titled ““A skill I wrote isn’t triggering when I expect it to.””The description field in its SKILL.md is the only signal Claude uses to decide relevance — sharpen it with an explicit trigger phrase describing when to use it. See Skills & Custom Commands.
“I keep getting interrupted by permission prompts.”
Section titled ““I keep getting interrupted by permission prompts.””That’s the system working as intended, not a bug — see Security & Client Data before reaching for --dangerously-skip-permissions as the fix. If prompts are genuinely just noisy for a specific safe tool, scope --allowedTools / allowed-tools more precisely instead of skipping all permission checks.
“Claude keeps writing code that doesn’t match our project’s conventions.”
Section titled ““Claude keeps writing code that doesn’t match our project’s conventions.””This usually means CLAUDE.md doesn’t actually state the convention explicitly, or states it too vaguely. See the shared CLAUDE.md template at Shared Conventions & CLAUDE.md.
“Which model should I even be using?”
Section titled ““Which model should I even be using?””See Model & Cost Guidance — short version: mid-tier for most day-to-day work, top-tier for genuinely hard or high-stakes tasks.
“Is it safe to point Claude Code at a client’s repo?”
Section titled ““Is it safe to point Claude Code at a client’s repo?””Read Security & Client Data in full before doing this on anything client-facing — short version: yes, with the same care you’d apply to any external tool touching client code.
“A session feels stuck or unresponsive.”
Section titled ““A session feels stuck or unresponsive.””Ctrl+C cancels the current action without killing the session. If Claude seems to be looping or the plan has clearly gone sideways, it’s usually faster to /clear or /rewind and re-prompt with a tighter instruction than to keep waiting it out.
“Claude used a lot of tokens for something that felt simple.”
Section titled ““Claude used a lot of tokens for something that felt simple.””Long-running sessions, large files pulled into context, and verbose tool output all add up. Check /context to see what’s actually filling the window, and see Model & Cost Guidance for ways to keep cost in check.
“Do I need an MCP server for everything?”
Section titled ““Do I need an MCP server for everything?””No — Claude Code’s built-in tools (file read/write, Bash, web search) cover most day-to-day work. Reach for MCP when you need Claude talking to something external it doesn’t already have a tool for. See MCP & Built-in Tools.
If something in this guide looks out of date, that’s expected — Claude Code moves fast, and this guide notes wherever a claim is likely to age. Check the official docs when something here seems off.