diff --git a/sdk/search/azure-search-documents/.github/skills/azure-search-documents/SKILL.md b/sdk/search/azure-search-documents/.github/skills/azure-search-documents/SKILL.md
index cb11558f91c5..a0982992a6dc 100644
--- a/sdk/search/azure-search-documents/.github/skills/azure-search-documents/SKILL.md
+++ b/sdk/search/azure-search-documents/.github/skills/azure-search-documents/SKILL.md
@@ -15,6 +15,7 @@ Use these sidecars for guidance that is required for release-ready changes but d
 
 - Release tracks, branch-dependent API surface, and CHANGELOG rules: `references/release.md`
 - Test commands, live recording workflow, and test authoring rules: `references/testing.md`
+- Sample authoring workflow, `.http` contract parity, and live-run setup: `references/samples.md`
 
 ## Environment setup
 
diff --git a/sdk/search/azure-search-documents/.github/skills/azure-search-documents/references/samples.md b/sdk/search/azure-search-documents/.github/skills/azure-search-documents/references/samples.md
new file mode 100644
index 000000000000..d9eb29a30ca7
--- /dev/null
+++ b/sdk/search/azure-search-documents/.github/skills/azure-search-documents/references/samples.md
@@ -0,0 +1,20 @@
+# azure-search-documents samples
+
+Samples run live against a search service. Create `Set-LiveSampleEnvironment.ps1` from the release resources to set the env vars the samples read, keeping secrets out of it. Run commands from `sdk/search/azure-search-documents/` using the `venv` alias from `SKILL.md`:
+
+```powershell
+cd sdk/search/azure-search-documents
+. .\.github\skills\azure-search-documents\scripts\Set-LiveSampleEnvironment.ps1
+venv python samples/sample_knowledge_source_file_preview.py
+```
+
+## Adding new samples
+
+This checklist is mandatory for every sample change. Do not add or update samples without checking each item below.
+
+1. Gather the three release inputs first: the sample requirements, `.http` calls, and the resources to run against. Ask before editing if any is missing or ambiguous.
+2. Name the file after the scenario from the requirements: `sample_<scenario>[_preview][_async].py`.
+3. Build the file body: a DESCRIPTION/USAGE docstring, `run_tag`-based resource names so reruns don't collide, and the demonstrated calls and their SDK imports wrapped in `# [START <name>]` / `# [END <name>]` tags with id = file name.
+4. Mirror each scenario's `.http` calls as the source of truth for the SDK flow: same operations (one SDK method per verb + route), call order, parameters, body fields, headers, and options, using SDK models instead of raw JSON; only swap in hotel data. If a starter value fails live, change only that value, not the call shape.
+5. Keep the full SDK flow visible inside the tags, down to credential and client construction; push setup and cleanup into `sample_utils.py` helpers. Helper-only data may use compact dicts; the demonstrated scenario may not.
+6. Mirror each sync sample in async; only async clients, `async with`, `await`, iteration, and cleanup differ.