We built the tool we wanted to hire.
Creative by Request is a creative studio. Tars is the AI advisor we built for ourselves first, used on our own work for months, and then opened to clients. This is how it got here, including the parts that did not work.
Most problems that arrive looking like creative problems are not. They are a strategy problem wearing a creative costume, or an operations problem that finally surfaced as a message nobody believes. A studio that only does one of those things will solve the visible half and leave the rest.
That is the reason this studio works across four disciplines rather than one, and it is also the reason we ended up building software. We kept wanting a second opinion that had actually read the strategy document before commenting on the campaign.
What was actually missing
General-purpose AI is good at producing something plausible. That is not the same as producing something right, and in creative work the gap between the two is the whole job. Ask a general model what your brand should say and it will answer from the average of everything it has read, which is a description of the middle of the market. The middle of the market is where brands go to be forgotten.
What was missing was not more fluency. It was a tool that argued from a specific company's own principles, and that would say so when a plan contradicted them.
Less, but better.
That line is one of our own operating beliefs, and it turned out to be the hardest one to build. A model will always give you more. Getting it to give you less, and to make the less count, took considerably longer than making it talk.
What we learned building it
Three things changed how we work, and all three came from being wrong first.
Measuring beats arguing. We were confident a larger model would give better advice, and we were ready to pay for it. Then we built a harness that ran real cases through the real prompt and measured the result. The larger model was three and a half times slower and no better. What actually improved the advice was rewriting the instructions and deleting a template it had been applying by rote. Being able to reject your own plan on evidence is worth more than being right early.
A model should not write its own memory. We built an automatic system that summarised what it learned about you after every conversation and carried it into the next one. It was elegant and it was wrong. A claim the tool made in one conversation became a stored fact that justified itself in the next. That whole mechanism is deleted. The standing note about you is now written by you, and nothing else can touch it.
Prompts ask; code decides. Told not to invent a company name it cannot verify, a model mostly complies. Mostly is not a standard you can sell. So the checking is done in code against real sources rather than requested politely in the instructions.
Where it runs
On our own hardware. Your strategy documents, your uploaded decks, your unreleased packaging and your voice do not leave the building, because sending a client's confidential material to a third-party API to save ourselves some engineering is not a trade we are willing to make on your behalf.
What you can use today
- Tars Public is open to anyone. Talk through a problem, upload a PDF for review, search the web with sources cited. Nothing is kept after the session ends.
- Tars Enterprise reads your own strategy, brand and operations documents and answers from them, with image review, voice, and an interview mode that helps you write those documents in the first place.
- Tars Vision is a free, browser-based tool for generative projection work. No account, no model, nothing to install.
The studio still does the work. Tars is what we built so that the work starts from what a company actually believes rather than from what everyone else has already said. If you are weighing a rebrand, a launch, or whether the thing you are about to build is worth building, that is the conversation we want.