Workflow · Production

Val Town's journey fast-following AI code assistants into Townie: from Copilot to Claude 3.5 Sonnet

The problem

Val Town users demanded a state-of-the-art LLM code generation experience from day one, but early implementations — a ChatGPT cosplay autocomplete and a function-calling Townie — were slow, inaccurate, and had poor feedback loops that made iterating on generated code difficult.

First attempt

The ChatGPT-powered autocomplete occasionally forgot its role and gave inaccurate completions, and the function-calling Townie hallucinated non-existent functions while its generic interface was ill-suited for iterating on code.

Workflow diagram · grounded in source
1
User describes app in Townie chat
trigger
“you can say, 'make me a ChatGPT clone with persistent thread history', and in about 30 seconds, you'll have a deployed app that does exactly that”
2
Claude generates fullstack app
ai_action
“Townie can generate a fullstack app, with a frontend, backend, and database, in minutes, and fully deployed”
3
Automatic error detection and fix offer
feedback_loop
“the LLM automatically detects errors and asks you if you'd like it to try to solve them. We detect server-side errors by polling our backend for 500 errors in your logs. We detect client-side errors in the iframe by prompting Townie to i…”
Reported outcome

Val Town shipped a Claude 3.5 Sonnet-powered Townie in August 2024 that can generate a fullstack app with frontend, backend, and database in minutes, fully deployed, and contributed a novel automatic error detection feature.

Reported metrics
Code generation timeminutes to generate just a couple hundred lines of code
App deployment timeabout 30 seconds
Reported stack
TownieCodeiumClaude 3.5 SonnetClaude ArtifactsChatGPTGPT-3.5
Source
https://blog.val.town/blog/fast-follow/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Val Town shipped a Claude 3.5 Sonnet-powered Townie in August 2024 that can generate a fullstack app with frontend, backend, and database in minutes, fully deployed, and contributed a novel automatic error detection f…

What tools did this team use?

Townie, Codeium, Claude 3.5 Sonnet, Claude Artifacts, ChatGPT, GPT-3.5.

What results were reported?

Code generation time: minutes to generate just a couple hundred lines of code; App deployment time: about 30 seconds (source-reported, not independently verified).

What failed first in this deployment?

The ChatGPT-powered autocomplete occasionally forgot its role and gave inaccurate completions, and the function-calling Townie hallucinated non-existent functions while its generic interface was ill-suited for iterati…

How is this workflow AI workflow structured?

User describes app in Townie chat → Claude generates fullstack app → Automatic error detection and fix offer.