diff --git a/README.md b/README.md
index 0953d5b..d2116b1 100644
--- a/README.md
+++ b/README.md
@@ -34,22 +34,35 @@ super-man ai "count lines in js files"  # AI bash one-liner via local LLM (advan
 ### AI mode (local LLM via ollama)
 
 `ai` generates a bash one-liner from plain English using a local [ollama](https://ollama.com)
-model. The model is configurable, so you can A/B different local LLMs:
+model. The default is `westenfelder/NL2SH`, a model purpose-built for
+natural-language-to-shell — in our 25-query benchmark it was the most accurate, the
+only one that handled hard multi-stage tasks. The model is configurable, so you can
+still A/B other local LLMs:
 
 ```bash
-super-man ai "find python files modified last week"          # default model
-super-man ai --model qwen2.5:3b "count lines in all js files"  # override model
-super-man ai --compare "list the 5 largest files"             # run several models side by side
+super-man ai "find python files modified last week"          # default model (NL2SH)
+super-man ai -q "list the 5 largest files"                   # quiet: print ONLY the command
+super-man ai --model qwen2.5-coder:1.5b "count lines in js"  # override model
+super-man ai --compare "list the 5 largest files"            # run several models side by side
+```
+
+The `-q` / `--quiet` / `--raw` flag prints only the generated command to stdout (errors
+go to stderr), so it composes in scripts and pipelines:
+
+```bash
+cmd=$(super-man ai -q "compress the logs folder into a tarball")
+echo "$cmd"            # tar -czf logs.tar.gz logs/
 ```
 
 | Env var | Purpose | Default |
 |---------|---------|---------|
-| `SUPERMAN_AI_MODEL` | Default model for `ai` | `qwen2.5-coder:1.5b` |
-| `SUPERMAN_AI_COMPARE_MODELS` | Space-separated list for `--compare` | `qwen2.5-coder:1.5b westenfelder/NL2SH` |
+| `SUPERMAN_AI_MODEL` | Default model for `ai` | `westenfelder/NL2SH` |
+| `SUPERMAN_AI_COMPARE_MODELS` | Space-separated list for `--compare` | `westenfelder/NL2SH qwen2.5-coder:1.5b` |
 
-General coder models (e.g. `qwen2.5-coder`) are prompt-wrapped to emit a single bare
-command; the purpose-built `westenfelder/NL2SH` model receives the raw query. Pull
-either with `ollama pull <model>`.
+The purpose-built `westenfelder/NL2SH` model receives the raw query; general coder
+models (e.g. `qwen2.5-coder`) are prompt-wrapped to emit a single bare command. Pull
+either with `ollama pull <model>`. See the
+[benchmark write-up](tests/performance/) for the full model comparison.
 
 `Esc` in the flag browser returns to the command list; `Esc` on the home screen
 exits.
diff --git a/super-man.sh b/super-man.sh
index 1e0d56e..083853a 100755
--- a/super-man.sh
+++ b/super-man.sh
@@ -41,7 +41,7 @@ show_help() {
   echo -e "  ${CYAN}Natural Language Queries:${NC}"
   echo -e "    ${GREEN}ask${NC} \"query\"           Ask in natural language (database)"
   echo -e "    ${GREEN}task${NC} \"description\"    Describe what you want to do (database)"
-  echo -e "    ${GREEN}ai${NC} \"complex query\"    AI bash one-liner via local LLM (advanced) 🤖"
+  echo -e "    ${GREEN}ai${NC} \"complex query\"    Text-to-bash one-liner, fully local (advanced) 🤖"
   echo ""
   echo -e "  ${CYAN}Browse & Explore:${NC}"
   echo -e "    ${GREEN}category${NC} NAME          Browse by category (files/text/network/system)"
@@ -76,19 +76,21 @@ show_help() {
   echo -e "  ${GREEN}✓${NC} jq (for natural language queries)"
   echo -e "  ${GREEN}✓${NC} fzf (for interactive mode only)"
   echo ""
-  echo -e "${BOLD}${YELLOW}AI MODE${NC} 🤖"
-  echo -e "  Generate bash one-liners from plain English via a local ollama model:"
+  echo -e "${BOLD}${YELLOW}AI MODE${NC} 🤖 (text-to-bash, fully local via ollama)"
   echo -e "    ${GREEN}super-man ai \"find python files modified last week\"${NC}"
-  echo -e "    ${GREEN}super-man ai --model qwen2.5:3b \"count lines in all js files\"${NC}"
-  echo -e "    ${GREEN}super-man ai --compare \"list the 5 largest files\"${NC}   # A/B models"
+  echo -e "    ${GREEN}super-man ai -q \"count lines in all js files\"${NC}      # print ONLY the command"
+  echo -e "    ${GREEN}super-man ai --compare \"list the 5 largest files\"${NC}  # A/B models"
+  echo ""
+  echo -e "  ${DIM}-q/--raw prints just the command — pipeable:${NC}"
+  echo -e "    ${GREEN}\$(super-man ai -q \"...\")${NC}   or   ${GREEN}super-man ai -q \"...\" | tee cmd.sh${NC}"
   echo ""
   echo -e "  Default model: ${CYAN}${SUPERMAN_AI_MODEL}${NC}"
   echo "  Configure via env vars:"
-  echo "    SUPERMAN_AI_MODEL           default model (e.g. westenfelder/NL2SH)"
+  echo "    SUPERMAN_AI_MODEL           default model (e.g. qwen2.5-coder:1.5b for speed)"
   echo "    SUPERMAN_AI_COMPARE_MODELS  space-separated list for --compare"
   echo ""
   echo -e "${YELLOW}Database:${NC} $DB_FILE"
-  echo -e "${YELLOW}Version:${NC} 2.2.0 (250+ commands, configurable local AI)"
+  echo -e "${YELLOW}Version:${NC} 2.3.0 (250+ commands, local text-to-bash AI)"
   echo ""
 }
 
@@ -319,8 +321,8 @@ mode_category() {
 # side with `--compare`. NL2SH is purpose-built for NL->shell so it receives
 # the raw query; general coder models get an instruction wrapper plus markdown
 # fence stripping so they too emit a single bare command.
-SUPERMAN_AI_MODEL="${SUPERMAN_AI_MODEL:-qwen2.5-coder:1.5b}"
-SUPERMAN_AI_COMPARE_MODELS="${SUPERMAN_AI_COMPARE_MODELS:-qwen2.5-coder:1.5b westenfelder/NL2SH}"
+SUPERMAN_AI_MODEL="${SUPERMAN_AI_MODEL:-westenfelder/NL2SH}"
+SUPERMAN_AI_COMPARE_MODELS="${SUPERMAN_AI_COMPARE_MODELS:-westenfelder/NL2SH qwen2.5-coder:1.5b}"
 
 # Build a model-appropriate prompt for a query.
 ai_build_prompt() {
@@ -374,11 +376,13 @@ ai_explain_base() {
 
 mode_ai() {
   local compare=false
+  local quiet=false
   local model="$SUPERMAN_AI_MODEL"
   local args=()
   while [ $# -gt 0 ]; do
     case "$1" in
       --compare|-c) compare=true ;;
+      --quiet|--raw|-q) quiet=true ;;
       --model|-m) shift; model="$1" ;;
       *) args+=("$1") ;;
     esac
@@ -387,27 +391,43 @@ mode_ai() {
   local user_query="${args[*]}"
 
   if [ -z "$user_query" ]; then
-    echo -e "${RED}Error: No query provided${NC}"
-    echo "Usage: super-man ai [--model NAME|--compare] \"your query\""
-    echo ""
-    echo "Examples:"
-    echo "  super-man ai \"find all python files modified in last week\""
-    echo "  super-man ai --model qwen2.5:3b \"count lines in all js files\""
-    echo "  super-man ai --compare \"list the 5 largest files\""
+    echo -e "${RED}Error: No query provided${NC}" >&2
+    echo "Usage: super-man ai [--model NAME] [--quiet|--compare] \"your query\"" >&2
+    echo "" >&2
+    echo "Examples:" >&2
+    echo "  super-man ai \"find all python files modified in last week\"" >&2
+    echo "  super-man ai -q \"count lines in all js files\"   # prints only the command" >&2
+    echo "  super-man ai --compare \"list the 5 largest files\"" >&2
     return 1
   fi
 
   # Check if ollama is available
   if ! command -v ollama &> /dev/null; then
-    echo -e "${RED}Error: Ollama is not installed${NC}"
-    echo ""
-    echo "Install Ollama:"
-    echo "  curl -fsSL https://ollama.com/install.sh | sh"
-    echo ""
-    echo "For now, try: super-man ask \"$user_query\""
+    echo -e "${RED}Error: Ollama is not installed${NC}" >&2
+    echo "" >&2
+    echo "Install Ollama:" >&2
+    echo "  curl -fsSL https://ollama.com/install.sh | sh" >&2
+    echo "" >&2
+    echo "For now, try: super-man ask \"$user_query\"" >&2
     return 1
   fi
 
+  # ── Quiet/raw mode: print ONLY the generated command (pipeable, scriptable) ─
+  if $quiet; then
+    if ! ai_model_installed "$model"; then
+      echo "Error: model '$model' not installed. Pull with: ollama pull $model" >&2
+      return 1
+    fi
+    local out
+    out=$(ai_generate "$model" "$user_query")
+    if [ -z "$out" ]; then
+      echo "Error: no response from $model" >&2
+      return 1
+    fi
+    printf '%s\n' "$out"
+    return 0
+  fi
+
   # ── Compare mode: run every configured model side by side ──────────────────
   if $compare; then
     echo ""