Ask HN: How do you manage bad artifacts?
2 comments
We have a similar setup, but for our whole business at display.dev
We have dedicated document lifecycle skills, folder setup and frontmatter fields in markdown to maintain the order.
We have a temp/ folder, which is for temporary files and does not get committed at all. We also have folders like foundations/ and scratch/ which the name says, former you can trust and latter not really. Inside top level folders we also have separate _wip/ folders for WIP stuff, which have not yet been promoted.
Markdown files have frontmatter whether it’s a draft, plan or it’s already implemented. Also last updated date is included for the agent to evaluate freshness.
In addition we also have scripts which validate the frontmatter fields and surface broken links before pushing the changes.
On top of all of that we use display.dev to share the artifacts with others in the team and get feedback.
Have you looked at LLM Wiki concept from Karpathy? It might have some additional ideas on how to better manage the knowledge base.
At Display.dev we’re also working on a solution to replace the need for Github repo. So you could manage the artifacts without knowing that it’s git underneath. So stay tuned.
We have dedicated document lifecycle skills, folder setup and frontmatter fields in markdown to maintain the order.
We have a temp/ folder, which is for temporary files and does not get committed at all. We also have folders like foundations/ and scratch/ which the name says, former you can trust and latter not really. Inside top level folders we also have separate _wip/ folders for WIP stuff, which have not yet been promoted.
Markdown files have frontmatter whether it’s a draft, plan or it’s already implemented. Also last updated date is included for the agent to evaluate freshness.
In addition we also have scripts which validate the frontmatter fields and surface broken links before pushing the changes.
On top of all of that we use display.dev to share the artifacts with others in the team and get feedback.
Have you looked at LLM Wiki concept from Karpathy? It might have some additional ideas on how to better manage the knowledge base.
At Display.dev we’re also working on a solution to replace the need for Github repo. So you could manage the artifacts without knowing that it’s git underneath. So stay tuned.
We use Cursor to work in here mostly on biz-dev, non-technical proejcts.
A lot of what is generated is ideation, and 75% or more is unused. But it gets stored as MD files regardless. So future agents use some of this as context for thinking about whatever the new session is exploring. Which sucks and means it gets more noisy over time.
Anyone have a good method for cleaning up old artifacts automatically that doesnt involve manually deleting MD files or text inside MD files?