Add beginner guide, LLM guardrails disclosure, fix bugs

- Add docs/GETTING-STARTED.md for non-technical users
- Add "Known Limitations" section to README about LLM safety guardrails
- Add FAQ entry about AI seeming evasive on trauma topics
- Fix setup.sh prompt_choice outputting to stdout instead of stderr
- Fix first session detection in CLAUDE.template.md (check for empty
  sessions folder before trying to reference previous sessions)
- Add guardrails acknowledgment instructions to CLAUDE.template.md
  (AI should be honest when hitting built-in limitations)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

README.md

@@ -35,6 +35,8 @@ This toolkit solves all of them.
 
 ## Quick Start
 
+**New to Claude Code or terminal-based AI tools?** See the [Getting Started Guide](docs/GETTING-STARTED.md) for step-by-step instructions.
+
 ### Prerequisites
 
 - [Claude Code](https://claude.ai/code) (recommended) or another LLM interface
@@ -223,6 +225,37 @@ Optional but recommended for shared computers. Setup script walks you through it
 
 ---
 
+## Known Limitations
+
+### LLM Safety Guardrails
+
+All major LLMs (Claude, GPT, etc.) have built-in safety filters. These can affect therapy conversations in ways you should know about:
+
+**Topics where the AI may seem hesitant or awkward:**
+- Sexual trauma and assault
+- Abuse (especially detailed descriptions)
+- Self-harm (specifics, not general discussion)
+- Violent experiences
+- Suicidal ideation (the AI will pivot to crisis resources)
+
+**What this looks like:**
+- The AI may give generic responses instead of engaging deeply
+- It may redirect the conversation or change the subject
+- It may add disclaimers or suggest professional help more than necessary
+- In some cases, it may refuse to continue a line of discussion
+
+**This is not the AI judging you.** It's hitting guardrails built into the model. The AI cannot override these filters.
+
+**What you can do:**
+- Frame topics in terms of feelings and impact rather than graphic details
+- If the AI redirects, try rephrasing: "I want to process how this affected me emotionally"
+- Understand that some trauma processing may require a human therapist who can hold space for the full story
+- Consider a local LLM with fewer guardrails (see SECURITY.md), though this trades off response quality
+
+This limitation is real, and we want you to know about it upfront rather than feel dismissed mid-session.
+
+---
+
 ## FAQ
 
 **Can I use this with ChatGPT instead of Claude?**
@@ -245,6 +278,10 @@ Yes. Edit CLAUDE.md directly, or purchase expansion packs for additional modalit
 
 This tool is designed for everyday emotional support. If you're dealing with severe symptoms, trauma, or crisis situations, please seek professional help. The AI will also recommend this when appropriate.
 
+**Why does the AI seem evasive when I talk about trauma?**
+
+LLMs have safety guardrails that can make them awkward around certain topics (abuse, sexual trauma, self-harm, violence). This isn't the AI judging you — it's hitting built-in filters it can't override. See "Known Limitations" above for more detail and workarounds.
+
 ---
 
 ## Contributing