👋 Hey, this is Artem and this is the second issue of The Blueprint. At WIQ, we spend each day forward deployed at Fortune 500 companies working through the hairy journey of AI transformation.
This newsletter is our way of sharing the lessons we learn with our friends on the same journey.
This week’s artifact: we turned the prompt behind a historic math proof into the checklist we run before every deployment. Steal it below 🧪 The Enterprise Prompt Checklist
Last week, GPT 5.6 Sol Ultra proved the Cycle Double Cover Conjecture, a graph theory problem that had been open for fifty years. OpenAI released the prompt they used to do it.
If I had to steal one line from the prompt that did it, it would be:
"Spend at least 8 hours on this before even thinking of returning or giving up."
The model finished in under an hour; the 8 hour line was written in case it didn't. Why is it interesting? The prompt is two pages with barely any math in it, and it reads like the brief a good manager writes for a team facing a very hard problem.
Here's my breakdown of how the prompt was written and how we're applying it to enterprise AI transformations.
The case for middle management (for AI agents)
Strip out the graph theory and the prompt is a manager’s job description. It puts a diverse team on the problem and keeps them from converging on one idea. It gives one agent the job of breaking every answer. It defines done before any work starts, and it says a best-effort summary does not count. None of that needs a PhD.
Here’s what the PDF is doing:
Define the task. The exact problem and the win condition, explicitly noting partial credit does not count.
Manage the search. The model is asked to run 64 agents from different approaches. The prompt withholds the favored approach so they don't all converge ("Do not tell most agents the currently favored approach"), keeps a registry of approach families, and marks dead ends as blocked.
Audit everything. Adversarial agents whose only job is to break every candidate proof.
Don't come back early. "Return only when a complete affirmative proof has been found and survives adversarial audit." This, in combination with the 8 hour requirement, was the most interesting to me. Perhaps models should ship with a persistence toggle next to the intelligence toggle.
Persistence > Intelligence
Our head of product recently recommended a book called Grit (he just had a kid, great parenting book, would highly recommend). The core idea is that perseverance is a much better predictor of success than intelligence.
That’s basically what the prompt is enforcing in the model. Sure Sol Ultra is plenty smart, but it’s prone to giving up (or rationalizing a dead end) early.
Nobody can hand-write this doc for every task
The models have been smart enough for twelve months. We just don't manage them well.
In May I asked Claude to redline an MSA due back to a customer the next morning. It did a great job: caught the usual things, cleaned up the language, looked airtight.
My lawyer called the next day about one buried clause, rewritten in a way that would have voided our corporate insurance. The model never knew our policy, the jurisdiction-specific language, or the side letter my lawyer and I had agreed six months earlier.
The same class of intelligence that proves fifty-year-old conjectures nearly voided my insurance. The model was the same in both stories. The only difference is that OpenAI’s researchers wrote two pages of instructions and I typed two sentences into a chat box.
Most enterprises are amateur chefs chopping a salad 🥗 with a katana.
Last week I broke down what we call a blueprint: watch how a job is actually done, then turn it into an agent that can do it.
OpenAI hand-wrote that document for one problem. Your operations run on hundreds, and no one can hand-write hundreds of briefs or keep them current every time a tariff moves a shipping route or IT swaps a tool.
The brief has to come from the work itself, and stay current as the work changes.
The checklist we run before every deployment
We made a checklist version of this for ourselves. Before we point an agent at any operations task, we run it: eight checks, each with the question we actually ask.
Steal it here: The Enterprise Prompt Checklist.
If you would rather we build one from your real workflow, grab time at getwiq.ai/demo.
The Blueprint. Stories and lessons on AI transformations from founders and operators making it happen.