Your material
PDFs, articles, feeds and notes: what the agent does with them and what to give it.
Material is what keeps the agent specific. Without it, the agent writes from your brief and general knowledge — perfectly usable for a point-of-view post, thin when the post needs a fact, a figure or something that happened this week. With it, posts quote your book, react to today's ruling, and cite your own article rather than someone else's.
Two kinds of material
The setup conversation asks about material twice, because there are two different things worth having:
- Things you know from experience — the "grounded" question. Real situations, patterns you have seen, mistakes people make, numbers you have watched play out. These become the war stories and the concrete examples. Anonymise freely; the agent never names a client.
- Documents and feeds — the content sources step. Files and links the agent reads every time it writes, so posts can draw on your published work and on what is happening in your field right now.
Source types
| Type | Good for | What the agent reads |
|---|---|---|
| A book you wrote, whitepapers, reports, long articles, slide decks. Up to 50 MB per file. | A random five-page window each time it writes — so a 200-page book yields months of different posts without repeating itself. | |
| RSS feed | Industry news, a regulator's announcements, a publication you follow, your own blog. | The ten most recent entries, headline and a short excerpt of each. Read fresh every time, so the agent sees this morning's news at this morning's slot. |
| Web page | One article, a landmark ruling, your "about" page, a service page whose framing you want reflected. | The page text, up to about 8,000 characters, read fresh each time. |
| Website | Your whole site, a resource library, a firm blog. | Its sitemap, sampling up to 50 pages with a short excerpt of each, within an overall cap so the agent gets breadth rather than one page in depth. |
| Text | Notes, talking points, a client FAQ, your standard answers, pasted posts or emails. Up to 50,000 characters. | All of it, every time. This is also where the head-start step stores what it read from your LinkedIn. |
The description field matters more than it looks
Every source has a description, and it is the first thing the agent reads about it. "IRS newsroom feed — react to enforcement announcements, ignore routine notices" produces very different posts from a feed with no description. Tell the agent what the source is for: what to pull from it, what to skip, how much weight to give it.
What to give a tax attorney's agent
- An RSS feed from the IRS newsroom or Tax Notes, described as "current enforcement news — comment as a practitioner, don't summarise".
- The PDF of your book or your longest client guide, described as "my own writing — quote freely, attribute to me".
- A Text source of the eight questions every new client asks and how you answer them.
- Your firm's "about" page, so the agent frames your practice the way you do.
How sources behave once they are in
- Read fresh each time. Feeds and pages are fetched when the agent sits down to write, never cached from setup. Update the page, and the next post sees it.
- Shared across your agents. A source is added once and attached to any agent that should use it. Remove it from one agent and it stays available to the others.
- Never shared across people. For agencies: material added for a client belongs to that client. It appears only in that client's agents, never in another client's picker or on your own Content Sources page, and it is deleted with the client.
- Optional. An agent with no sources is normal — many voice-first agents run that way for months.
- Facts are checked against them first. When a draft states something checkable, the claims check reads it against your material before anything else, and treats what you provided as the authority on your own experience. See How your agent writes a post.
- A source that stops loading is flagged, not ignored. Feeds die, pages move, a PDF link can expire. Every time the agent writes, it records whether each source loaded; one that failed shows a warning next to its name (on the Content Sources page, the agent's console and the client hub) with the reason. The agent keeps writing from the rest of your material in the meantime, and the weekly question below knows not to count on it.
When material runs thin: the weekly question
Real material runs out. A brief with four war stories carries a month or two of posts, then the agent starts circling the same ground. Rather than let that happen, or make things up, the agent asks: once a week (or once a month, or never; set it on the agent's page under "Ask me for more material"), it emails you one specific question about your actual experience, chosen by looking at what your brief already holds, what the last posts covered, and what you have been asked before. "You mentioned the audit-reconsideration route. What is the first thing you tell a client whose reconsideration was denied?" is the sort of question you get, not "tell me more about yourself".
- A sentence or two back is plenty. The answer link opens a single box; specifics beat polish. Your words are merged into the brief the same way a "refine" in the console is, and anything new is marked fresh for two weeks so the agent builds a post around it within days rather than filing it away.
- Skip in one tap. The email's "Skip this one" link needs no login. Skips are signal: a skipped topic is asked about less, and one skipped twice is retired.
- Never more than one a period, never at a bad time. Questions go out on weekday mornings in the agent's time zone, never while the agent is already waiting on your client for something, never in the agent's first week (a fresh brief has not run thin yet), and not at all when there is nothing specific worth asking.
- Stop any time. Every question carries a "stop these questions" link, and the setting is on the agent's page. "Ask me one now" on the same panel sends a question immediately.
- For agencies: the question goes to whoever runs the client's agent (the assigned account manager, else the owner). Tick "Also email the client the question directly" to send the client a white-label copy in your agency's name with their own answer link; you are emailed when they answer, and their words land in the brief with everything else. See Emails you and your clients receive.
Your material is only ever used for your posts. Files are stored privately, sent to the writing model only as part of your own agent's prompt, and never used to train shared models or to write for anyone else. Delete a source and it is gone from every future post. See What we store, and what we don't.