diff --git a/docs/documentation-database.md b/docs/documentation-database.md
index fd72edcc1c..566703ad1b 100644
--- a/docs/documentation-database.md
+++ b/docs/documentation-database.md
@@ -9,6 +9,7 @@ This is a **read-only, prebuilt** database — CoGo never creates or migrates it
 - Installed path: `context.getDatabasePath("documentation.db")` (`Environment.DOC_DB` in `common/.../utils/Environment.java`), i.e. the app's private `databases/` dir.
 - Bundled as an asset and extracted on install/update by `BundledAssetsInstaller` / `SplitAssetsInstaller`.
 - **Debug override:** if `/sdcard/Download/documentation.db` exists and is newer than the installed copy, `WebServer` and `ToolTipManager` swap to it at request time (timestamp-compared per request, not just at startup) — a fast way to test a new database on-device without reinstalling. `WebServer`'s debug logging and experiment flags are also file-flag-gated under `/sdcard/Download/` (`CodeOnTheGo.webserver.debug`, `CodeOnTheGo.exp`, `CodeOnTheGo.webserver.cs0`).
+- **Don't trust a local copy's on-disk schema or row content as ground truth without checking freshness first.** Any manually downloaded or debug-override copy is independent of git history — a stale one can have a different schema (e.g. missing `UNIQUE(path)` or `templateId`) or be missing rows that already exist in the current, maintained database. A stale copy caused a real near-miss in ADFA-5088: a SQL script validated against it would have silently overwritten curated production tooltip content for several tags. Diff or re-download before authoring SQL against a local copy's state, not just before shipping it.
 
 ## Schema
 
@@ -91,6 +92,14 @@ Schema changes and data edits happen **outside this repo**, in `OfflineDocumenta
 - Tooltip uniqueness is `(categoryId, tag)`; `TooltipButtons.uri` values are validated there against `Content.path` (post `?query`/`#fragment` stripping) before being allowed into the database.
 - Every edit made through the tool updates `LastChange` for the affected documentation set, which is how `DatabaseVersionResolver`'s debug logging can say what build of the docs is loaded.
 
+### Writing one-off SQL scripts against this database
+
+Some tickets (e.g. ADFA-5088) ship a one-off `.sql` script under `docs/docdb/` for a `docdb-studio` maintainer to run against the real database, rather than editing it directly through the tool. Gotchas found writing those scripts:
+
+- **Keep each `.system` line simple.** The sqlite3 CLI's `.system` dot-command can hit a content-dependent shell-parsing failure when a line chains multiple operators (`;`, `&&`, `||`, parentheses) — it reproduces for some input strings and not others, so it won't necessarily show up in a quick test. Stick to one plain `command | pipe > file` per `.system` line.
+- **`.bail on` is required for `BEGIN`/`COMMIT` to actually mean atomic.** Without it, a mid-script SQL error prints to stderr but the script *keeps going* — including reaching the final `COMMIT`, which then persists whatever succeeded before the error (verified empirically, not just documented behavior). `.bail` also can't see `.system` shell failures directly, so a failed or empty Brotli payload (which leaves its target file missing or zero-length) needs its own check: insert its `READFILE()` into a throwaway `CREATE TEMP TABLE` guarded by `NOT NULL CHECK (length(content) > 0)` immediately before the real `Content` insert, turning that failure into a real SQL error `.bail` will catch. See `docs/docdb/ADFA-5088-preference-tooltips.sql` for the working pattern.
+- **Don't write Brotli payloads to bare `/tmp/*.br` filenames.** A fixed, guessable name directly under world-writable `/tmp` lets another local user pre-plant a symlink or race the write/read pair between the `.system echo | brotli` write and the `READFILE()` read (CWE-377). Create an owner-only working directory instead — `rm -rf` it, then `mkdir -m 700` it (the mode is set atomically at creation, with no window where it's briefly world-accessible) — write every payload under that directory, and remove it again before `COMMIT`. See the same script for the working pattern.
+
 ## Known rough edges
 
 - A Tier 3 link that points off-device is a bug in the *content*, not the code — the web server and webview will happily follow it. If you see one while working in this area, it's a data problem to report upstream, not a `WebServer` bug to fix here.