diff --git a/skills/basecamp/SKILL.md b/skills/basecamp/SKILL.md
index 94a5fb71..c5e9f1eb 100644
--- a/skills/basecamp/SKILL.md
+++ b/skills/basecamp/SKILL.md
@@ -91,12 +91,13 @@ Full CLI coverage: 189 tracked in-scope endpoints across todos, cards, messages,
 2. **Parse URLs first** with `basecamp url parse "<url>"` to extract IDs
 3. **Comments are flat** - reply to parent recording, not to comments
 4. **Check context** via `.basecamp/config.json` before assuming project
-5. **Content fields accept Markdown and @mentions** — message body and comment content accept Markdown syntax; the CLI converts to HTML automatically. Use Markdown formatting (lists, bold, links, code blocks, tables) for rich content. Four mention syntaxes are available (prefer deterministic for agents):
+5. **Content fields accept Markdown, and most accept @mentions** — the CLI converts these rich-text fields from Markdown to HTML: message bodies, document bodies, comment content, todo descriptions, card bodies, schedule entry descriptions, upload descriptions, check-in answers and notes. Two rich-text fields are sent as written, so give them HTML: todolist descriptions and gauge needle descriptions. Chat is different again: `chat post` sends plain text unless you pass `--content-type text/html` or the line carries a mention. Use Markdown formatting (lists, bold, links, code blocks, tables) for rich content. @mentions resolve in message bodies, comment content, card bodies, schedule descriptions and chat lines — not in todo descriptions, documents, uploads, check-ins or notes. Four mention syntaxes are available (prefer deterministic for agents):
    - **`[@Name](mention:SGID)`** — zero API calls, embeds SGID directly (preferred for agents)
    - **`[@Name](person:ID)`** — one API call, resolves person ID to SGID via pingable set
    - **`@sgid:VALUE`** — inline SGID embed for pipeline composability
    - **`@Name` / `@First.Last`** — fuzzy name resolution (may be ambiguous)
-   For todos, documents, and cards, content is sent as-is — use plain text or HTML directly.
+
+   Raw HTML is also accepted, but it is all-or-nothing per field: a tag the CLI detects as HTML (`<p>`, `<ul>`, `<strong>`, `<a>`, `<img>`, `<table>` and the other common formatting tags) outside a backtick code span or backtick fence (a `~~~` fence does not hide it) skips Markdown conversion for the whole field, so any Markdown alongside it — `![alt](/local/path)` included — is sent literally. The HTML itself goes through as written, except that an empty separator paragraph is inserted between directly adjacent `<p>` blocks so they render with spacing; a local path in a raw `<img src>` is still uploaded and replaced with an attachment in every converted field except notes, which take no attachments. Titles (a todo's content argument, card and message titles) are plain text and never converted.
 
    **Table boundary:** GFM tables round-trip: they render in message/comment
    bodies, display converts them back to pipe tables, and the TUI in-place