diff --git a/skills/hotdata/SKILL.md b/skills/hotdata/SKILL.md
index ed17605..58063b9 100644
--- a/skills/hotdata/SKILL.md
+++ b/skills/hotdata/SKILL.md
@@ -89,6 +89,12 @@ Returns workspaces with `public_id`, `name`, `active`, `favorite`, `provision_st
 
 **File formats:** `databases tables load` accepts **csv**, **json**, and **parquet** (local `--file`, remote `--url`, or a pre-staged `--upload-id`). The format is read from the file's extension; `--format csv|json|parquet` overrides it, and an unrecognised extension is left to the server to resolve from the bytes. **json is read in any shape** — an array of objects, a pretty-printed document, or one object per line: `--file`/`--url` reshape it to newline-delimited json before upload, so no `jq` step is needed. Every row must be an object, and the reshape is byte-faithful per row — key order and number digits are preserved. Two inputs are refused locally, before anything uploads: a row that is not an object, and a source with no rows at all (`[]`, or an empty file). Note the **load itself** reads json numbers as f64, whatever the shape: an integer wider than i64/u64 or a decimal past ~17 significant digits lands rounded — use parquet, or a string column, where the exact digits matter.
 
+**Loading — three checks before you trust a load:**
+
+- **A csv must be comma-separated.** A `;`- or tab-separated file loads **with no error** into a single column named `id;name` — the load has no delimiter option, so nothing reports this. After loading a csv you did not write yourself, run `databases tables show` and check the column list before you query or report success. (A `.tsv` extension currently fails with a 500.)
+- **"json in any shape" is the CLI's doing, not the API's.** `--file`/`--url` reshape an array or a pretty-printed document into newline-delimited json locally. The load API and the SDKs accept **newline-delimited json only** (`format: "json"`), so code that moves off the CLI must convert the file itself.
+- **Two json sources are refused before anything uploads — do not retry them unchanged.** `one object per row` means a row is a scalar or a list: reshape the file so every row is an object. `carries no json rows` means the source is `[]` or empty: there is nothing to load.
+
 **Active database:** `hotdata databases use <id>` saves the active database to config. `databases tables list`/`load`/`remove`, `databases queries`/`results`, and all `databases context` commands default to the active database; pass **`--database <id>`** to override per-command. (`databases tables show` instead takes a fully-qualified `catalog.schema.table`.)
 
 **Always select databases by id** (`dbid...`, from `databases list`). Display names and catalog aliases are not unique — several databases can share a name, and a fork answers to the same catalog as its source — so name-based selection is ambiguous.