Full-stack dev building AI-integrated web apps with Next.js & TypeScript.
Psychology MA | Intelligent Systems MSc.
Currently exploring how AI can make community feedback actually useful.
Location
London, UK
Education
MSc Intelligent Systems (UCL), MA Psychology (Oxford)
Civic technologist building AI tools that help systems explain themselves. Clew Suite builder, NACM 2026 speaker. I write about hackathons, product judgment, agent readiness, and learning in public.
Full-stack dev building AI-integrated web apps with Next.js & TypeScript.
Psychology MA | Intelligent Systems MSc.
Currently exploring how AI can make community feedback actually useful.
Location
London, UK
Education
MSc Intelligent Systems (UCL), MA Psychology (Oxford)
The insight isn't the adaptive questions it's that you built a survey that learns what it doesn't know. Most data collection tools optimize for completion. You optimized for coverage. That's the difference between counting answers and understanding a problem.
Full-stack dev building AI-integrated web apps with Next.js & TypeScript.
Psychology MA | Intelligent Systems MSc.
Currently exploring how AI can make community feedback actually useful.
Location
London, UK
Education
MSc Intelligent Systems (UCL), MA Psychology (Oxford)
That's a much better way of putting it than i did in the article. Coverage over completion. Traditionally survey would be optimised for 'did they finish', here it's optimised for 'did we learn something new'
Adaptive surveys are a great AI use case because the value is obvious and the failure mode is gentle - a slightly-off follow-up question is low-stakes, but a sharper one materially lifts response quality. Static surveys leave so much on the table by asking everyone the same thing; branching on the actual answer is the unlock.
The bit I'd watch as it scales: keep the question-generation grounded so it doesn't wander into leading or off-topic questions (which quietly bias your data). A light validation pass on generated questions pays off. The "automate the tedious 80%, guard the part that matters" pattern again - same thesis behind Moonshift (prompt to a shipped SaaS on your own GitHub+Vercel). Cool product; is the adaptation per-response live, or does it learn across all respondents over time? (Moonshift's first run's free if useful.)
Really like the “smart follow-ups from emerging gaps” angle — that’s closer to how good human interviewers work than most static forms.
A couple of research-design notes that might help anyone trying this pattern:
Guardrails on follow-ups — Cap how many adaptive probes you fire per respondent, or fatigue will erase the depth you gained. I’ve seen adaptive flows work best when the core instrument stays short (5–8 items) and AI only digs on ambiguous or extreme answers.
Comparability — Purely adaptive paths make cross-wave comparison harder. Logging which probes fired (and why) helps later analysis stay honest.
Bias check — If the model prefers certain themes for follow-ups, you can accidentally amplify early patterns. Sampling a few full transcripts weekly is worth it.
Would love to hear how you evaluate whether a follow-up actually improved insight quality vs. just adding more text.
The dual-context prompt design is smart — feeding both individual answers and aggregate gaps. Have you noticed the AI questions getting too narrow as the dataset grows, or does the gap detection keep them balanced?
Full-stack dev building AI-integrated web apps with Next.js & TypeScript.
Psychology MA | Intelligent Systems MSc.
Currently exploring how AI can make community feedback actually useful.
Location
London, UK
Education
MSc Intelligent Systems (UCL), MA Psychology (Oxford)
So it tracks which response category has the least coverage and prompts "we have very few responses about X." So as dominant themes get well-covered, the AI gets nudged towards what's underrepresented rather than drilling further into what everyone's already talking about. Haven't tested past ~100 responses yet though.
We're a place where coders share, stay up-to-date and grow their careers.
Fascinating
Thanks Ben!
Great build with so many applications!
Thanks! It's deployed in a community feedback context now. But the would work anywhere qualitative data is collected at scale
The insight isn't the adaptive questions it's that you built a survey that learns what it doesn't know. Most data collection tools optimize for completion. You optimized for coverage. That's the difference between counting answers and understanding a problem.
That's a much better way of putting it than i did in the article. Coverage over completion. Traditionally survey would be optimised for 'did they finish', here it's optimised for 'did we learn something new'
Adaptive surveys are a great AI use case because the value is obvious and the failure mode is gentle - a slightly-off follow-up question is low-stakes, but a sharper one materially lifts response quality. Static surveys leave so much on the table by asking everyone the same thing; branching on the actual answer is the unlock.
The bit I'd watch as it scales: keep the question-generation grounded so it doesn't wander into leading or off-topic questions (which quietly bias your data). A light validation pass on generated questions pays off. The "automate the tedious 80%, guard the part that matters" pattern again - same thesis behind Moonshift (prompt to a shipped SaaS on your own GitHub+Vercel). Cool product; is the adaptation per-response live, or does it learn across all respondents over time? (Moonshift's first run's free if useful.)
Really like the “smart follow-ups from emerging gaps” angle — that’s closer to how good human interviewers work than most static forms.
A couple of research-design notes that might help anyone trying this pattern:
Would love to hear how you evaluate whether a follow-up actually improved insight quality vs. just adding more text.
The dual-context prompt design is smart — feeding both individual answers and aggregate gaps. Have you noticed the AI questions getting too narrow as the dataset grows, or does the gap detection keep them balanced?
So it tracks which response category has the least coverage and prompts "we have very few responses about X." So as dominant themes get well-covered, the AI gets nudged towards what's underrepresented rather than drilling further into what everyone's already talking about. Haven't tested past ~100 responses yet though.