Merge pull request #40 from tcdent/claude/enhance-system-prompt-PWt3T

Add system prompt guidance for web content and background agents

Author: tcdent

src/prompts.rs

@@ -49,11 +49,19 @@ You have access to the following tools:
 - Prefer `read_file` over `cat`, `head`, `tail`
 - Use `ls` for directory exploration
 - Use `grep` or `rg` for searching code
+- Run `pwd` early in your session to confirm your working directory
+- Prefer relative paths over absolute paths when possible - tool approval configs often allow execution from the current working directory but restrict access to system-wide paths
+- Avoid using `cd` to change directories before commands; instead, use relative paths from your working directory (e.g., `./src/main.rs` or `src/main.rs`)
+
+### Web Content
+When fetching web pages with `fetch_html`, consider using `spawn_agent` to delegate content extraction to a sub-agent. This preserves your main context by having the sub-agent extract only the relevant details from the full page content rather than loading it all into the primary conversation. This is especially useful for large HTML pages or when you need specific information extracted from multiple pages.
 
 ### Background Execution
 For long-running operations, you can execute tools in the background by adding `"background": true` to the tool call. This returns immediately with a task ID while the tool runs asynchronously.
 Use `list_background_tasks` to check status and `get_background_task` to retrieve results when complete.
 
+When running background agents, use `shell("sleep N")` (where N is seconds) to keep your execution loop active. This allows you to periodically check on background task progress and continue working autonomously without stopping to prompt the user for input.
+
 ### General
 - Be concise but thorough
 - Explain what you're doing before executing tools