Alex Bristow

nxk

A skill that lets a coding agent ask a crowd of simulated people about a decision while it works. It shows who answers differently, and the agent keeps asking until it knows what to change.

Year
2026
Type
Open source skill
Source
github.com/albri/nxk ↗
Choosing a library's opening hours, one question at a time.

A crowd an agent can ask

A coding agent often makes decisions that affect people: the wording of a sign-up page, a price, a default setting or, in this page's made-up example, a library's opening hours. nxk gives it a crowd to ask first. It's a skill, a file of instructions that a coding agent loads when a task calls for it. The agent puts a question to a crowd of simulated people, sees who answers differently and decides what to look at next, without leaving the task.

The people come from PersonaGen, my other project. It makes synthetic people for the UK and the US from published statistics, so a crowd of 1,000 has the country's real mix of ages, incomes and households, and each person's job, money and home fit together.

Say the agent is helping to choose a library's opening hours. It first asks each person how likely they are to visit, and retired people are far more likely than people in full-time work. Every later answer is weighted by that chance, which puts mornings ahead. Full-time workers still lean towards evenings, and a follow-up finds the evening they'd use. The agent puts 3 hours on weekday mornings and 2 on Thursday evening.

What the model alone misses

Ask a language model on its own and you get one answer for everyone, a blend of all the people a decision might affect. The useful part is usually the difference: who wants evenings, who would pay more, who misreads the wording. A model answering as one particular person can show that, and I wanted to know how well.

I tried around 60 general-purpose models. Who I asked and how I framed the question mattered more than how capable the model was. Written answers sounded like one assistant with different names, and a single vote hid whether a person leaned 51% or 95% towards an answer. So nxk has the agent ask a concrete question and keep each person's probability for every option.

To be worth asking mid-task, a study also had to be quick and cheap. With general-purpose models it meant hundreds of requests at once, some taking 30 seconds, with retries when a structured answer came back malformed, and cost about 50p. Decision models are built for this: they return a probability for every option and write no text. nxk recommends Jev from TypeSafe, and OpenAI's Decisions API and Cloudflare's open Clef models work the same way. In one recorded run with Jev, 100 people answered 2 questions in 5 seconds for about 1p of model input cost.

Built onSKILL.md, PersonaGen, Jev, or another answering model

Following the answers

The library is one decision. The bigger use is a longer run, where each answer decides the next question. The skill tells the agent to start where the answer is unknown, drop directions that go nowhere and stop when another answer wouldn't change the work.

I gave an agent a made-up budgeting app, Tally, with a deliberately generic landing page, and let it follow the answers. In about 35 minutes and for about 20p it changed who the app is for, what it does, what it costs and what it says. It also left 2 leads that only a small group cared about. Run again on 160 different people, every finding held.

That is where nxk earns its place. Nobody has published a survey about this page. The finding everything else rested on was who the app is for, and one blended answer would have hidden it.

Checked against published surveys

Before trusting it on questions like these, I checked it on ones that already have answers: 26 questions from 15 published UK and US surveys, using an earlier version of PersonaGen. Each question went to 1,000 simulated people, and the comparison was the same model with no person attached. The model alone can't show how groups differ at all. With nxk, 12 of the 16 biggest gaps between groups point the same way as the survey, 9 of them clearly. Most come out at about a third of their real size, and the gaps by money come out too big. nxk's answers were also closer to the surveys, 9.7 percentage points off per answer on average against 14.0.

The results are modest, and they may flatter both runs. Published surveys are likely to be in the models' training data, so some of what looks like insight could be recall. It's a check that nxk points the right way, not proof that it's accurate, and recreating a survey that already exists is the least useful thing it does.

I turned the misses into checks. Some came from asking the wrong audience. Others were the model's own biases, with people sounding too positive or too cautious. A model can weigh a new offer, but it can't know private facts that a profile doesn't hold. nxk now tells the agent to check that its audience can answer the question, to flag results that look suspiciously uniform, and to report the group, the sample size and any failed answers with the result.

From a research app to a skill

nxk started as a full research app, with guided setup, question checks, raw responses and reports. I spent a long time on it, and it was close to launch. It was good at setting up survey-style studies and asking one-off questions. But the value I kept finding was in the follow-ups, where one answer raised a question I hadn't planned, and in an app each of those meant setting up another study.

I wanted nxk inside the work, not in another app where each question meant 5 minutes of setting up a study. A coding agent already has the context of its task and can ask again whenever it needs to. So I stopped the app and made nxk a skill. It's a bet on the agent: that it will reach for nxk when a decision depends on people, and keep asking until the answers stop changing what it does.

It's open source under the MIT licence. Installing it takes one command, a free PersonaGen key and a model to answer with.