Where AI marketing actually breaks
We build the brief that stops a model inventing claims, the gate that blocks what it gets wrong, and the measurement that tells us what reached anyone.
Email usThe draft wanted to say: Our products are compliant in 20 states.
The compliance file lists 14. Nobody had checked.
So we build the brief from your documents — and the gate blocks any number they do not support.
checks run: claim_verified · length
Why this fails
Three causes, in the order they cost you money.
No brief.
A prompt is not a brief. Give a model a topic and it will return a confident paragraph containing a number nobody can source. We write briefs with an explicit permitted-claims list. If a claim is not on that list, the draft cannot use it — and if the draft needs a number we do not have, the agent stops and asks instead of estimating.
No gate.
Most teams check output by reading it, which works until the week you are busy. We check in code first: length, banned phrases, repeated hooks, unverified claims, personal data. A blocked draft does not reach a human for approval — it goes back to the writer with reasons, twice at most, and then it fails.
No measurement.
Without a fixed baseline every post looks like learning. We keep an eval set of known-good and known-bad drafts and score every prompt change against it. Without that, you cannot tell an improvement from a coin flip. We have run this long enough to know we could not.
How it works
Five stages. Two of them are code, not judgment.
- Brief — a signal becomes a brief with a permitted-claims list and an angle.
- Draft — the writer produces a carousel and a caption to the channel's limits.
- Gate — deterministic checks first, then a model pass on claims and tone. A block is final; the writer cannot override it. (code)
- Publish — a human approves until the pipeline has earned the right not to need one.
- Measure — engagement, visits and replies are recorded against the post that caused them.
Two gates hold the whole thing up. Warden blocks anything from leaving. The eval gate blocks any change to the pipeline from being promoted before it has been scored. Both are checks in a repository, not policies in a document.
How an engagement runs
Audit — week 1.
We read what you publish now, what claims you can actually defend, and where your briefs would come from. You get a written gap list, including the parts we think you should not automate.
Build — weeks 2 to 4.
We write the brief format, the voice spec and the gate rules, then run the first ten drafts through them. You get a pipeline you can operate without us, and the fixtures that prove it works.
Operate — ongoing.
One change at a time, each scored against the eval set before it is promoted. You get a weekly review that separates what we observed from what we think it means.
Proof
Tropism is new. There are no client logos on this page, and we are not going to invent any.
- 11 agents
- 17 fixtures
- 2 gates
What exists is our own pipeline, run in public: eleven agents, one of them operational, seventeen eval fixtures and two gates, and a registry that reports an unbuilt agent as unbuilt rather than as ready. This page was written by hand, not by the pipeline. When it is written by the pipeline, we will say so.
If you would rather see the machinery than the marketing, ask for the eval set.
Contact
[email protected] — we answer within a day. If your problem is the wrong shape for us, we will say so rather than sell you a pilot.