diff --git a/README.md b/README.md
index 01a9093..a947069 100644
--- a/README.md
+++ b/README.md
@@ -15,7 +15,11 @@
   <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-blue.svg" alt="license: MIT"></a>
 </p>
 
-<!-- Demo GIF goes here once recorded: `brew install vhs && vhs docs/demo.tape` -->
+<p align="center">
+  <img src="https://raw.githubusercontent.com/devcodes9/agsearch/main/docs/demo.gif"
+       alt="Typing a misspelled half-remembered phrase; the session you meant is row one"
+       width="100%">
+</p>
 
 ## Why not just `/resume`?
 
@@ -168,7 +172,7 @@ opened the session *asking for*, so it's weighted hardest; **length normalizatio
 sprawling transcript that mentions your term once no longer outranks a short, on-point session.
 Near-ties are then nudged by **recency** and by how often you've actually resumed that session
 (read off agsearch's own resume log). Badge `N/T` = query terms matched. A word that barely exists
-anywhere (a typo like `alibrry`) falls back to subsequence matching. No modes or flags.
+anywhere (a typo like `conection`) falls back to subsequence matching. No modes or flags.
 
 **Honest limit:** this is lexical, not semantic. Ranking beats an unordered `rg` dump, but it
 is not a guarantee that row one is the session you meant. Because transcripts are large,
diff --git a/agsearch b/agsearch
index 6578cdb..04339bb 100755
--- a/agsearch
+++ b/agsearch
@@ -1041,7 +1041,7 @@ def build_sessions(lines):
         fh.write("\n".join(out))
 
 
-# Common words that add noise, not signal, to a search ("migration OF alibrary").
+# Common words that add noise, not signal, to a search ("migration OF the database").
 _STOP = {"of", "the", "a", "an", "to", "for", "in", "on", "and", "or", "is", "it", "this",
          "that", "with", "from", "by", "at", "as", "be", "are", "was", "were", "how", "do",
          "i", "my", "me", "we", "you", "about", "into", "using", "use", "some", "any"}
@@ -1219,7 +1219,7 @@ def rank_sessions(rows, qterms, usage=None, now=None):
     length normalization, so a sprawling session no longer outranks a short exact match, and
     what a session was *opened to do* outweighs a passing mention buried in it. The result is
     then nudged by recency and how often you've resumed that session. A term that exists almost
-    nowhere as text (a typo like 'alibrry') falls back to subsequence matching.
+    nowhere as text (a typo like 'conection') falls back to subsequence matching.
     """
     usage = usage or {}
     n = len(rows) or 1
diff --git a/docs/demo.gif b/docs/demo.gif
new file mode 100644
index 0000000..c26cd4c
Binary files /dev/null and b/docs/demo.gif differ
diff --git a/docs/demo.tape b/docs/demo.tape
deleted file mode 100644
index 0d38c2a..0000000
--- a/docs/demo.tape
+++ /dev/null
@@ -1,42 +0,0 @@
-# VHS tape for the README demo GIF.  https://github.com/charmbracelet/vhs
-#
-#   brew install vhs
-#   vhs docs/demo.tape        # writes docs/demo.gif
-#
-# The tape is committed rather than just the GIF on purpose: re-recording after
-# a UI change is `vhs docs/demo.tape`, not getting the take right again.
-#
-# Keep it under ~15 seconds. The demo's whole job is to show that row one is
-# the session you meant — not to tour the flags.
-
-Output docs/demo.gif
-
-Set Shell "zsh"
-Set FontSize 16
-Set Width 1200
-Set Height 640
-Set Padding 20
-Set Theme "Catppuccin Mocha"
-Set TypingSpeed 60ms
-
-Hide
-Type "clear" Enter
-Show
-
-Sleep 700ms
-Type "agsearch"
-Sleep 400ms
-Enter
-
-# Cold list settles, then type the thing you half-remember.
-Sleep 2s
-Type "stripe tax"
-Sleep 2500ms
-
-# The point of the whole product: the preview pane already shows the matched
-# line, highlighted, before spending a token.
-Sleep 1500ms
-
-# Enter resumes — into Claude Code or Codex, whichever the session came from.
-Enter
-Sleep 3s