diff --git a/products/ai_observability/skills/exploring-llm-traces/SKILL.md b/products/ai_observability/skills/exploring-llm-traces/SKILL.md
index 020bc004db68..b7835fd08e41 100644
--- a/products/ai_observability/skills/exploring-llm-traces/SKILL.md
+++ b/products/ai_observability/skills/exploring-llm-traces/SKILL.md
@@ -246,7 +246,7 @@ posthog:query-llm-traces-list
 
 For more complex SQL patterns, read these references:
 
-- [Single trace retrieval](./references/example-llm-trace.md.j2) — fetches a single trace by ID with all events and properties (renders the `TraceQuery` HogQL)
+- [Single trace retrieval](./references/example-llm-trace.md) — fetches a single trace by ID with all events and properties (renders the `TraceQuery` HogQL)
 - [Traces list with aggregated metrics](./references/example-llm-traces-list.md) — two-phase query: find trace IDs first, then fetch aggregated latency, tokens, costs, and error counts
 
 ## Parsing large trace results
diff --git a/products/posthog_ai/scripts/build_skills.py b/products/posthog_ai/scripts/build_skills.py
index 2f55c184aa25..147d51ebc1d1 100755
--- a/products/posthog_ai/scripts/build_skills.py
+++ b/products/posthog_ai/scripts/build_skills.py
@@ -266,6 +266,62 @@ def _assert_text_file(file_path: Path) -> None:
         )
 
 
+_MARKDOWN_LINK_RE = re.compile(r"\]\(([^)\s]+)\)")
+
+
+def _check_reference_links(entry: Path, repo_root: Path) -> list[str]:
+    """Find markdown links to references/ or scripts/ files that will not exist in the built bundle.
+
+    A skill ships its SKILL.md, references/, and scripts/ files, and the build strips the .j2 suffix
+    from every rendered template. So a link to `references/x.md` resolves when the source holds either
+    `references/x.md` or `references/x.md.j2`, but a link to `references/x.md.j2` never resolves,
+    because the bundle only has the stripped `references/x.md`.
+    """
+    skill_dir = entry.parent
+    bundle: set[str] = set()
+    for subdir_name in sorted(_ALLOWED_SUBDIRS):
+        subdir = skill_dir / subdir_name
+        if not subdir.is_dir():
+            continue
+        for root, dirs, filenames in os.walk(subdir):
+            dirs[:] = sorted(dirs)
+            for filename in sorted(filenames):
+                # The build strips .j2 when it renders a template, so record the shipped path.
+                rel = str((Path(root) / filename).relative_to(skill_dir))
+                bundle.add(rel.removesuffix(".j2"))
+
+    # Only the SKILL.md entry point is scanned. Reference files link to each other with paths
+    # relative to the skill root, which do not resolve from inside references/, and their code
+    # snippets hold `](...)` fragments that are not links.
+    errors: list[str] = []
+    text = entry.read_text()
+    source_label = str(entry.relative_to(repo_root))
+    for match in _MARKDOWN_LINK_RE.finditer(text):
+        target = match.group(1).split("#", 1)[0].split("?", 1)[0]
+        if not target or "://" in target or target.startswith("mailto:"):
+            continue
+        resolved = (skill_dir / target).resolve()
+        try:
+            rel_to_skill = resolved.relative_to(skill_dir.resolve())
+        except ValueError:
+            continue  # Link points outside the skill; not part of the bundle.
+        if not rel_to_skill.parts or rel_to_skill.parts[0] not in _ALLOWED_SUBDIRS:
+            continue  # Only references/ and scripts/ files ship in the bundle.
+        if resolved.is_dir():
+            continue  # A directory ships through its files.
+        if str(rel_to_skill) not in bundle:
+            line, _col = _line_col(text, match.start(1))
+            hint = (
+                "Link to the built '.md' path, not the '.md.j2' template."
+                if target.endswith(".j2")
+                else "No file of that name ships in the skill."
+            )
+            errors.append(
+                f"Broken reference link in {source_label}:{line}: '{target}' does not resolve to a bundled file. {hint}"
+            )
+    return errors
+
+
 class SkillFrontmatter(BaseModel):
     name: str
     description: str = Field(max_length=_MAX_SKILL_DESCRIPTION_LENGTH)
@@ -612,6 +668,9 @@ def lint_all(self) -> bool:
         for skill in skills:
             lint_files = self._collect_lint_files(skill)
 
+            if skill.depth == 1:
+                errors.extend(_check_reference_links(skill.source_file, self.repo_root))
+
             for file_path in lint_files:
                 source_label = str(file_path.relative_to(self.repo_root))
 
diff --git a/products/posthog_ai/scripts/test/test_build_skills.py b/products/posthog_ai/scripts/test/test_build_skills.py
index 9776d59a2403..8211f1b2b566 100644
--- a/products/posthog_ai/scripts/test/test_build_skills.py
+++ b/products/posthog_ai/scripts/test/test_build_skills.py
@@ -504,6 +504,30 @@ def test_lint_all_catches_bad_jinja2_in_subdirectory(tmp_path: Path) -> None:
     assert builder.lint_all() is False
 
 
+@pytest.mark.parametrize(
+    "link_target,expected",
+    [
+        ("references/payload.md.j2", False),
+        ("references/payload.md", True),
+        ("references/missing.md", False),
+    ],
+)
+def test_lint_all_checks_reference_links_against_the_bundle(tmp_path: Path, link_target: str, expected: bool) -> None:
+    skill_dir = tmp_path / "products" / "alpha" / "skills" / "linker"
+    skill_dir.mkdir(parents=True)
+    (skill_dir / "SKILL.md").write_text(f"---\nname: linker\ndescription: D\n---\nSee [payload]({link_target}).\n")
+    refs = skill_dir / "references"
+    refs.mkdir()
+    (refs / "payload.md.j2").write_text("# {{ 'rendered' }}\n")
+
+    builder = SkillBuilder(
+        repo_root=tmp_path,
+        products_dir=tmp_path / "products",
+        output_dir=tmp_path / "output",
+    )
+    assert builder.lint_all() is expected
+
+
 def test_lint_all_catches_duplicate_skill_names(tmp_path: Path) -> None:
     for product in ("alpha", "beta"):
         skill_dir = tmp_path / "products" / product / "skills" / "same-name"