How a brief gets built

A weekly brief looks like a short piece of writing. Most of the work behind it is throwing things away and refusing to make things up. Here's the whole pipeline, honestly, including the parts that fail.

1. Collect, from public sources only

Each week we pull the recent posts and their comments from the Instagram and TikTok accounts on your tracker, through licensed third-party data providers. Everything we read is what a logged-out visitor could see. We never log in on anyone's behalf, never ask for your credentials, never post or follow or message, and never touch a private account.

2. Throw most of it away

The first real step is deletion. A busy comment section is mostly emoji, bots, giveaway spam, tagged friends, and one-word praise. None of it is signal, and all of it is expensive to reason about, so it goes before any model sees it. Depending on the niche this removes the clear majority of what we collect. What survives is the sentences where somebody actually said something.

3. Classify what survives

Each surviving comment gets sorted into what it's doing: question, complaint, purchase intent, praise, comparison. This is the step that turns a pile of text into something countable — forty people asking the same unanswered question is a fact about your market, but only once you can count it.

Every model response here is schema-validated. If it comes back malformed, it gets one repair attempt, and then it fails loudly. It never gets waved through. An unvalidated classification is worse than no classification, because it looks exactly like a real one downstream.

4. Consolidate the week

Individual comments are anecdotes. The brief needs patterns, so we consolidate across every tracked account: what repeated, what changed against last week, which posts overperformed and what their comment sections have in common. A theme that appears once is noise. A theme that appears forty times, up from six last week, is the thing you should film.

5. Write it, with receipts attached

Only then does anything get written. The brief is generated against a strict schema in which every claim must carry the specific comments it came from. Evidence isn't decoration bolted on afterwards — a claim that can't cite its comments is structurally invalid and never reaches you.

This is why you can click into any line of a brief and read the raw comments underneath it. Not because we added a nice drill-down feature, but because the claim cannot exist without them.

6. Admit it when the data is thin

Some weeks a tracked account posts twice. Sometimes a provider returns partial data. The tempting move — the one that makes the product look better and makes it useless — is to write the same confident brief anyway.

Instead the brief is annotated: this week's sample was small, treat it accordingly. Partial data gets labelled as partial. It is never quietly topped up with plausible invention. A brief you can trust in a thin week is worth more than one that never admits weakness.

What we forget on purpose

Commenter identities never land in our database. At ingestion the username is replaced with a salted one-way hash whose only job is to stop us counting the same spammer twice. The evidence you read shows the comment, its like count, and which post it appeared under — never who wrote it.

This costs us features. We can't build follower-level targeting, we can't tell you which specific person to DM, and we can't answer a request to find everything one person ever said, because there is nothing to search. That's the intended outcome. The people in those comment sections didn't sign up for our product, and a market signal doesn't require knowing who they are.

Why any of this matters

It would be much easier to feed a few thousand comments to a language model and print whatever comes out. It would read well. It would be confidently wrong on a schedule, and you'd have no way to catch it — which is precisely the problem with acting on it.

Every constraint above exists to make the output falsifiable. You can check our work. That's the whole product.