docs(AGENTS): add sequential-thinking MCP and caveman skill requirements

- Require sequentialthinking MCP for all planning/problem-solving
- Add caveman mode for compressed user-facing communication
- Clarify karpathy-guidelines usage for task execution quality

AGENTS.md

@@ -33,6 +33,27 @@ Always use these skills when working with this repository:
 - **nix-flakes** - For reproducible builds, flake management, and devShell operations
 - **karpathy-guidelines** - For high-quality task execution best practices
 
+### Required MCP Server
+
+- **sequentialthinking** — MUST be used for ANY planning, problem-solving, or decision-making.
+  Invoke `sequentialthinking` MCP before:
+  - Designing implementation approaches
+  - Debugging non-obvious issues
+  - Evaluating trade-offs between approaches
+  - Breaking down complex tasks
+  - When you're uncertain about the best path forward
+
+  Do NOT skip deliberative thinking. Surface reasoning through sequentialthinking before acting.
+
+### Communication: Caveman Mode
+
+Use **caveman** skill for all user-facing communication. Responses should be:
+- Ultra-compressed — cut token waste, keep technical substance
+- Direct — no filler, no hedging, no "I'd recommend"
+- Action-oriented — state what was done and what's next
+
+Trigger: `skill: "caveman"` before responding to the user.
+
 ### Blueprint Framework
 
 This project uses `numtide/blueprint` which: