A 30-day plan, then a weekly read of your own numbers.

Generic AI writing plateaus fast, because it never finds out what happened next. The copilot closes that loop.

It starts by asking: who is this for, what do you sell, which networks matter, what does a win look like. From those answers it builds a 30-day plan and pre-loads it as drafts, each one already pointed at your connected accounts — not a blank calendar and a wish. The first week lands in Review so you approve it; the rest sits on the calendar so a month of posts does not arrive as a wall.

Before anything ships, each draft gets a predicted score per platform with the reasons behind it: hook strength, length against that network’s truncation point, CTA presence, posting time. Each reason comes with a fix you can apply in one click.

Then it watches. A weekly job reads this workspace’s own published results, derives what it can, and renders those findings into the drafting prompt as a labelled evidence block. Three things keep that honest. It is framed to the model as evidence rather than instruction — apply it where it does not fight the brief, the platform or the brand voice, and ignore any line that does. A quiet week produces nothing: a workspace with four published posts has nothing statistically honest to learn from, so the job records why it skipped instead of inventing a finding. And it is per workspace, so one client’s numbers never shape another client’s copy.

To be exact about what this is: a weekly evidence block computed from your own posts and placed in the prompt. It is not a fine-tuned model, not continuous training, and nothing learned in your workspace ever crosses into anyone else’s.