How your agent writes a post
Interview, brief, drafts, and a separate claims check — the pipeline behind every post.
You answer a handful of questions once. From then on, at each scheduled time, a post appears under your name that sounds like you and says something worth saying. This is what happens in between.
1. Your brief
Everything you told the agent during setup — who you are, what you speak on, who you write for, how you sound, what you would never say, what is off-limits, your posting style — is kept as one standing brief. It is the agent's instructions, written in your words, and you can read and edit it at any time from the agent's console. Each answer carries a small light showing whether it is rich enough to write from; a thin answer is the most common reason for a generic post.
2. Your material, read fresh
When it is time to write, the agent fetches your material as it is right now: the latest entries of each feed, the current text of each page, a random five-page window of each PDF. Nothing is cached from setup, so a post reacting to news reacts to this morning's news.
3. Your recent posts, and what you thought of them
The agent also looks at your most recent posts — full text, and any image concept — with your thumbs and notes attached. Posts you liked are examples to keep sounding like; posts you didn't, with your reason, are examples to steer away from. It is also told not to repeat a topic or an opening it used recently. This is why variety and voice both improve over the first few weeks without you doing anything except reacting.
4. The draft
With brief, material and history in hand, the writing model produces one post for the platform in question, within that platform's character limit (the tightest one, if the agent posts to several — see Platforms and limits). If it runs long, it is asked to shorten it, up to twice. If images are on, the model also proposes a one-sentence image concept in the style you chose.
5. The claims check
Writing and checking are two separate jobs. A second reader — a reviewer on the same tier of model as the writer, briefed as a senior peer in your field — reads the finished draft and pulls out every statement of fact: figures, dates, rules, rulings, names, "studies show". Each one is checked against your brief and material first, and against primary sources on the public record where your material cannot settle it. It is told what expert errors look like — a true rule stated with the wrong scope, two adjacent tests blended into one, a term of art swapped for its neighbour — and, just as firmly, what not to touch: your opinions, your own experience, a simplification that is still correct.
Anything wrong is corrected with the smallest edit that fixes it, in your voice, and the corrected draft is read again. If the second read is clean, the post carries a "Checked · 1 correction" badge you can click to see what was said, what it was changed to, and why. If after a couple of rounds the reviewer still cannot vouch for a passage, the post is held: it is never queued, never shown to you as a draft, and never published; the agent writes a fresh one instead. Nothing reaches the outbox, or you, unchecked. This is the mechanism that keeps a wrong figure or a misstated rule from ever appearing under your name.
6. The outbox
Every checked post goes straight into the outbox — the three posts from setup, posts written by the look-ahead email, posts you asked for — and the next one in line goes out at the next slot. A thumbs-up moves a post to the front; a thumbs-down takes it out. If the outbox is empty when a slot arrives, the agent writes, checks and publishes on the spot. Either way you never have to approve anything; the outbox exists so you can look, not so you must.
7. The image
If your agent uses images, the concept from step 4 is rendered as a square image in your chosen style, uploaded to private storage, and attached to the post. It is produced in the background so the text is never waiting on the picture: in the outbox you may briefly see "Image on its way". If the image cannot be produced in time, the post goes out as text on platforms that allow it; on image-only platforms the slot is skipped rather than posting a half-finished post under your name.
8. Publishing
At the scheduled minute, the post and its image are sent to the platform through our publishing partner, which holds the official integrations. Each target account is published independently — if X is down, LinkedIn still goes out. A failed attempt is retried three times over the next 45 minutes; a post that still cannot be published is marked "Failed to post" where you can see it, and nothing is ever published twice.
Where you come in
Nowhere, by default. When you want to: read the look-ahead email on Sunday, thumbs-up or thumbs-down a post in the outbox, drop a note about what to change, pull a post you would rather not run, pause the agent for a holiday. Each of those takes seconds and none of them is required. See Feedback that changes future posts.