diff --git a/src/screen-manager.ts b/src/screen-manager.ts
index 9d9c4a61..ea20a2a4 100644
--- a/src/screen-manager.ts
+++ b/src/screen-manager.ts
@@ -108,9 +108,12 @@ export class ScreenManager extends EventEmitter {
     const screenName = `claudeman-${sessionId.slice(0, 8)}`;
 
     // Create screen in detached mode with the appropriate command
+    // Set CLAUDEMAN_SCREEN=1 so Claude sessions know they're running in Claudeman
+    // This helps prevent Claude from attempting to kill its own screen session
+    const envVars = `CLAUDEMAN_SCREEN=1 CLAUDEMAN_SESSION_ID=${sessionId} CLAUDEMAN_SCREEN_NAME=${screenName}`;
     const cmd = mode === 'claude'
-      ? 'claude --dangerously-skip-permissions'
-      : '$SHELL';
+      ? `${envVars} claude --dangerously-skip-permissions`
+      : `${envVars} $SHELL`;
 
     try {
       // Start screen in detached mode
diff --git a/src/session.ts b/src/session.ts
index f85b1f47..f3ccf4c8 100644
--- a/src/session.ts
+++ b/src/session.ts
@@ -551,7 +551,13 @@ export class Session extends EventEmitter {
         cols: 120,
         rows: 40,
         cwd: this.workingDir,
-        env: { ...process.env, TERM: 'xterm-256color' },
+        env: {
+          ...process.env,
+          TERM: 'xterm-256color',
+          // Inform Claude it's running within Claudeman (helps prevent self-termination)
+          CLAUDEMAN_SCREEN: '1',
+          CLAUDEMAN_SESSION_ID: this.id,
+        },
       });
     }
 
@@ -707,7 +713,12 @@ export class Session extends EventEmitter {
         cols: 120,
         rows: 40,
         cwd: this.workingDir,
-        env: { ...process.env, TERM: 'xterm-256color' },
+        env: {
+          ...process.env,
+          TERM: 'xterm-256color',
+          CLAUDEMAN_SCREEN: '1',
+          CLAUDEMAN_SESSION_ID: this.id,
+        },
       });
     }
 
@@ -803,7 +814,13 @@ export class Session extends EventEmitter {
           cols: 120,
           rows: 40,
           cwd: this.workingDir,
-          env: { ...process.env, TERM: 'xterm-256color' },
+          env: {
+            ...process.env,
+            TERM: 'xterm-256color',
+            // Inform Claude it's running within Claudeman
+            CLAUDEMAN_SCREEN: '1',
+            CLAUDEMAN_SESSION_ID: this.id,
+          },
         });
 
         this._pid = this.ptyProcess.pid;
diff --git a/src/templates/claude-md.ts b/src/templates/claude-md.ts
index 9b442b82..573fbe8a 100644
--- a/src/templates/claude-md.ts
+++ b/src/templates/claude-md.ts
@@ -20,8 +20,17 @@ export function generateClaudeMd(caseName: string, description: string = '', tem
 
   return `# CLAUDE.md - Project Configuration
 
-## Project Overview
+## Setup
+Copy these files to your new project:
+- \`CLAUDE.md\` → project root
+- \`.claude/settings.json\` → \`.claude/settings.json\`
 
+Then update the Project Overview section below.
+
+---
+
+## Project Overview
+<!-- Update this section with project-specific details -->
 - **Project Name**: ${caseName}
 - **Description**: ${description || 'A new project'}
 - **Tech Stack**: [TECHNOLOGIES_USED]
@@ -29,6 +38,17 @@ export function generateClaudeMd(caseName: string, description: string = '', tem
 
 ---
 
+## Claudeman Environment
+
+This session is managed by **Claudeman** and runs within a GNU Screen session.
+
+**Important**: Check for \`CLAUDEMAN_SCREEN=1\` environment variable to confirm.
+- Do NOT attempt to kill your own screen session
+- The session persists across disconnects - your work is safe
+- Token usage, costs, and background tasks are tracked externally
+
+---
+
 ## Work Principles
 
 ### Autonomy
@@ -47,6 +67,52 @@ Full permissions granted. Act decisively without asking - read, write, edit, exe
 ### Thinking
 Extended thinking is enabled. Use deep reasoning for complex architectural decisions, difficult bugs, and multi-file changes.
 
+### Task Tracking (TodoWrite)
+**ALWAYS use TodoWrite** to track tasks. This is non-negotiable for anything beyond trivial single-step work.
+
+**When to use TodoWrite:**
+- Multi-step tasks (3+ steps)
+- Bug fixes requiring investigation
+- Feature implementations
+- Any work where progress tracking helps
+- When the user provides multiple requests
+
+**How to use it:**
+1. **Before starting**: Break down the work into discrete todos
+2. **During work**: Mark each todo \`in_progress\` before starting, \`completed\` when done
+3. **One at a time**: Only ONE todo should be \`in_progress\` at any moment
+4. **Immediately**: Mark todos complete the moment they're done - don't batch
+
+**Why this matters:**
+- Gives the user visibility into your progress
+- Prevents forgetting tasks mid-work
+- Creates accountability checkpoints
+- Makes complex work manageable
+
+**Example workflow:**
+\`\`\`
+User: "Add user authentication with JWT"
+
+→ TodoWrite:
+  - [ ] Research existing auth patterns in codebase
+  - [ ] Implement JWT token generation
+  - [ ] Add login endpoint
+  - [ ] Add token validation middleware
+  - [ ] Add protected route example
+  - [ ] Write tests
+
+→ Mark "Research existing auth patterns" as in_progress
+→ Do the research
+→ Mark as completed, mark next as in_progress
+→ Continue until all done
+\`\`\`
+
+**Anti-patterns to avoid:**
+- Starting work without creating todos first
+- Having multiple todos \`in_progress\` simultaneously
+- Batching completions at the end
+- Skipping TodoWrite for "simple" multi-step tasks
+
 ---
 
 ## When to Use Agents
@@ -100,8 +166,8 @@ Planning mode flow: read-only exploration → create plan → get approval → e
 Ralph loops enable persistent, autonomous work on large tasks. When active, you continue iterating until completion criteria are met or the loop is cancelled.
 
 ### Starting a Ralph Loop
-- Start: \`/ralph-loop\`
-- Cancel: \`/cancel-ralph\`
+- Start: \`/ralph-loop:ralph-loop\`
+- Cancel: \`/ralph-loop:cancel-ralph\`
 - Help: \`/ralph-loop:help\`
 
 ### Time-Aware Loops
@@ -146,6 +212,27 @@ echo "Elapsed: $ELAPSED_HOURS hours"
 - Dependency updates
 - Type safety improvements
 
+**Example time-aware prompt:**
+\`\`\`
+"Optimize the API endpoints for the next 4 hours. Focus on performance first,
+then code quality. Minimum runtime: 4 hours."
+Completion phrase: <promise>TIME_COMPLETE</promise>
+\`\`\`
+
+**Time-aware loop behavior:**
+\`\`\`
+[Start loop, record timestamp]
+[Complete primary optimization tasks - 2 hours elapsed]
+[Check time: 2/4 hours - NOT done yet]
+[Self-generate: "Add caching to database queries"]
+[Self-generate: "Optimize N+1 queries"]
+[Self-generate: "Add request batching"]
+[Continue working... 4.5 hours elapsed]
+[Check time: 4.5/4 hours - minimum reached]
+[All tasks complete, tests pass]
+<promise>TIME_COMPLETE</promise>
+\`\`\`
+
 ### How You Know You're in a Ralph Loop
 
 The user started the loop with a prompt containing:
@@ -239,6 +326,44 @@ WHEN ALL TASKS DONE:
 5. Only then output completion phrase
 \`\`\`
 
+### Example: How to Think During Ralph Loop
+
+**Original prompt**: "Add CRUD endpoints for todos with validation"
+
+**Your approach**:
+\`\`\`
+Task breakdown:
+- [ ] GET /todos (list)
+- [ ] POST /todos (create with validation)
+- [ ] GET /todos/:id (single)
+- [ ] PUT /todos/:id (update with validation)
+- [ ] DELETE /todos/:id
+- [ ] Tests for all endpoints
+
+Starting with GET /todos...
+[implement]
+[test - passes]
+[commit: "feat(todos): add GET /todos endpoint"]
+[update session log]
+
+Moving to POST /todos...
+[implement]
+[test - fails: validation not working]
+[fix validation]
+[test - passes]
+[commit: "feat(todos): add POST /todos with validation"]
+[update session log]
+
+...continue until all done...
+
+Final verification:
+[npm test - all pass]
+[npm run lint - clean]
+[npm run build - succeeds]
+
+<promise>COMPLETE</promise>
+\`\`\`
+
 ### When to NOT Output Completion Phrase
 
 - Tests are failing (even one)