How to use Claude to generate APIs with OctopusGarden
The author reports successful trials with mostly standard CRUD and REST API apps, while describing the generated code as messy and noting unresolved hardening and maintenance issues.
news.ycombinator.com
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Using [OctopusGarden setup], generate [CRUD or REST API application] from [specification]. Build it in Docker and validate against [holdout scenarios] withheld from the generating agent. Score scenario satisfaction and feed failures into another iteration. Report results and remaining limitations.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source author
- foundatron
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- The author reports successful trials with mostly standard CRUD and REST API apps, while describing the generated code as messy and noting unresolved hardening and maintenance issues.
- Human involvement
- The author tried the generated apps and assessed maintainability; human code review is explicitly absent from the factory loop.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Submitted specifications for mostly standard CRUD and REST API apps.
- Generated application code and built it in Docker.
- Validated builds against scenarios withheld from the generating agent.
OctopusGarden is the author's open-source implementation of a software factory pattern. They described it as automating connections between workflows they already used in Claude Code. A specification enters the factory, which generates code, builds it in Docker and validates it against holdout scenarios that the generating agent has not seen.
Validation uses probabilistic satisfaction scoring with an LLM judge. Failures feed back into a convergence loop that keeps iterating. The author tried this with mostly standard CRUD and REST API applications and reported that it worked. They had not tested HTML or JavaScript applications, and did not use the factory to build OctopusGarden itself.
The account also sets limits on that result. The generated applications worked, but their code was described as messy and difficult to maintain. The factory did not yet provide the complete linting, static checking and testing the author wanted. A Claude outage prevented smoke tests and new app generation, highlighting dependence on the model service. The author also identified unresolved security hardening and compliance concerns.
Who it's for: developers exploring specification-driven API generation who can assess the generated code and the factory's stated limitations.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Hacker News · foundatron
Frequently asked questions
What kinds of applications were tried?
The author tried mostly standard CRUD and REST API applications and said they worked. They had not tried HTML or JavaScript applications.
Was generated code reviewed by a human in the loop?
The factory pattern has no human code review in its convergence loop. The author separately assessed the results and described the code as messy and difficult to maintain.
More details
Remaining steps
- Scored satisfaction and fed failures back into the convergence loop.
About this account
The factory loops through generation, builds, hidden validation and feedback without human code review in the loop.
Platform: Hacker News. Attribution: https://news.ycombinator.com/item?id=47226107. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: foundatron on Hacker News
Source published
Added to Agents In Use
Source checked
Tools and topics
OctopusGarden · Claude · Docker · LLM judge
API generation · Docker · Validation
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.