๐Ÿ Claude Impact Lab London ยท 30 July 2026 ยท solo build

Three hours and ten minutes

I got to the venue early and walked the streets around it. Two corners away were two chalkboards, both full of local life, and I don't think either one is online. Then a woman spent twenty minutes at the museum's coffee cart telling me what actually goes on around there. I'd have missed all of it with my headphones in, which is what I'd normally have had in.

The brief

โ€œHow can AI help communities thrive? Instead of centralising opportunities, how can it expose them locally and match people to what's needed?โ€

The event brief, theme three

Theme three asked how AI could surface local opportunity and match people to it. I'd spent the hour before that looking at two examples of it, chalked up on a wall and invisible to anyone not standing in front of them.

The plan, and what it became

I planned it the night before and de-risked everything I could: proved the model path live, pinned the request shape, wrote a smoke test to run in the first ten minutes. The build order was a ladder, the chat matcher first, and nothing else until that held.

On the day one integrator session ran parallel Claude Code agent streams, each in its own git worktree, each merged and verified before it counted. The ladder held early, and the scope kept growing all evening.

  1. 15:30Smoke test green โ€” live model call confirmed, and four streams cut immediately.
  2. 16:20Live end to end โ€” real entries on the board and the matcher answering live, with a 40-check eval battery already forcing prompt rewrites.
  3. 16:45Demand detection โ€” working, and a year-on-year check measured 86% churn in listed opportunities.
  4. 17:40Snap โ€” photo or PDF into structured entries, contact details stripped back to roles, demoed live.
  5. 18:10Freeze โ€” demo board locked, though Journey, the any-place sketcher and a site audit all landed after it.
  6. 18:40Full run โ€” one sentence about baking returned a course, a stall venue, a cafรฉ and a ยฃ2,000 fund.
3h10the whole build window, solo
28changes, each stamped with a verify command at a commit
40automated checks the matcher prompt had to keep passing as it changed
86%of listed opportunities changed within a year, measured on the day

What I'd do differently

I think the factor I'd change is how many features I built. There was a point in the evening where I was expanding the scope because I could, and not because it needed it. The build order was meant to stop that, since I'd decided the night before that nothing started until the chat matcher was answering live, and then the next item only once that was proven. However it didn't stop me adding to it, and so once the chat matcher was working, I just kept going. I think three or four ideas polished properly would have made a better demo than twenty-eight things that mostly work.

It was my first hackathon and I probably underestimated what doing it solo really meant. I think a lot of the scope problem happened because there was nobody sat there to tell me no. I'm signing up for more hackathons, and next time I'd want to be in a team, either roping in a few people I know or joining one on the day.

The build, frozen