run · platform
people and AI, working together by design. own the capability.
Every skill, on the model that runs it best.
The Third Vector Platform runs your skills on several models and shows which one does each job best. It tests every new version against examples your experts have approved, and feeds what your team says about the output back into the skill, so it gets sharper with every run.
Models come and go. Your skills stay, and the Platform shows you which model to run them on. In early access today, by invitation, with the teams we build skills for.
Run it. Test it. Improve it.
A skill earns its keep once your team relies on it. From then on, three questions matter: which model should run it, does it still get the important cases right, and what has your team learned that the skill has not.
models
The best model for each skill
Run the same skill on several models, several times each, and compare them on one scorecard per model: which follows your rules, which asks for human judgement where it should, and which fails. Choose on evidence, and check again when a new model ships.
gold sets
Examples your experts sign off
A gold set is the handful of cases a skill has to get right, with the answers your experts approve. Your team keeps it on the Platform, and every new version of a skill, and every new model, is tested against it before anyone relies on it.
feedback
Feedback that reaches the skill
When someone on your team marks an output as off, their note goes back to the skill it came from. The next version is drafted from those notes and tested against the gold set, so the skill improves with use.
Designed to run on any model.
The skills we build run on Claude today, and they are portable by design: on another provider's model, through the proxy or gateway you already use, or against a model you host yourself. Results still differ from one model to the next, which is exactly what the Platform measures.
Built on Agent Skills, the open standard developed by Anthropic and supported by tools from OpenAI, GitHub, Google and Mistral. Your skills move with you if you change AI provider.
early access
By invitation, with the teams we build skills for.
Today the Platform runs the skills we build with clients in Applied AI. If your team already has skills of its own and wants to see how they hold up across models, tell us about them.
how we build skills →Judgement in the flow of the work.
Agentic AI compresses the analytical work: the research, the modelling, the drafting. The best outcomes come when people make their calls while that work is happening. The Third Vector Judgement Framework names the six calls that stay with them, and every skill we build is designed to draw each one out.
01
Intent
A frontier model runs with whatever intent it is given, at full speed. Why this work, why now, and what it needs to prove. Setting that is yours.
02
Key assumptions
Every case rests on a few load-bearing assumptions. Naming them, and choosing which to test first, is a human call.
03
How ideas are articulated
The way an idea is put into words shapes what the analysis returns. Sharper in, sharper out.
04
How factors are weighted
AI will happily list twelve factors that matter. Deciding that two of them decide the case is judgement.
05
Gaps in context
A model only knows what it has been given. Spotting what is missing from the room is on the people in it.
06
The gut check
When the analysis says yes and something still feels off, that signal counts. Good method makes room for it rather than overriding it.
the judgement layer
How the Platform puts the framework into practice.
A skill stops and asks at the points where a person has to decide, and the scorecard shows how often each model asks for human judgement where it should. Your team's answers, and their notes on the output, go back into the skill, so the next version asks better questions.
Human judgement, AI leverage, on a loop.
You set intent. AI analyses. You judge. Every pass compounds, and the capability stays with you.
01
You set the intent
Set the objective: why this work, why now, and what it needs to prove. That call is yours.
02
AI does the analysis
It researches, drafts, and models, sometimes interviewing you as it works.
03
You own the judgement
Weigh what matters, challenge the output, spot the gaps, and make the call.
A diagram of the Third Vector working loop: six stages orbiting the Third Vector mark in a continuous cycle. Human sets intent, AI runs research and drafting, AI presents output, human challenges it, AI refines, human judges. Each pass of the loop compounds into reusable building blocks.
How we keep quality high.
AI is probabilistic, not deterministic.
Run the same skill twice and you get different drafts. That is why the Platform runs each skill several times before it shows you anything, and why judgement stays in the loop.
Quality in, quality out.
The output is only as sharp as the input you give it.
AI makes mistakes, the way people do.
It can misread context, overstate confidence, or get a fact wrong. The gold set and the method are built to catch that.
Our Terms say the same thing more formally: skills are thinking tools, provided as is, and no outcome is guaranteed.
Find out which model runs your skills best.
Book a call and tell us about the skills your team relies on. We will show you how the Platform would run them.
book a call →