• system prompts are retrieved from supabase
• system prompts can be reviewed, added, and activated via admin ui
• temperature and other settings can be managed through admin ui
• save actual model variant used (from response)

• update supabase to store returned model version, settings
• build admin ui to manage active models, settings

Strategy to keep article storage on free tier: delete every x days and/or limit request size?

• Use Promise.all to send requests concurrently
• Set up supabase RPC for compound upsert

• When users view article details, they answer brief questions before viewing AI analysis
• How does this make you feel? (Angry, sad, hopeful, curious)
• After responding/skipping, they can see AI analysis compared to survey response percentages

• Add tooltips explaining language and politics
• Add a tagline noting which model provided the analysis

• Accept/reject individual components of analysis
• Choose a reason why rejected (?)

• Draft project history, goals, methodology, roadmap, links to more resources
• Integrate via Notion or MDX
• Add reporting on approved vs. rejected analyses by models (reach: maybe by news source, by topic?)

• Admin can log in via Supabase auth
• Admin can view all articles — WON'T DO
• Admin can "approve" AI-generated analysis of new articles
• Public users can view only approved articles — WON'T DO

• Users can submit a suggested topic to track (send to email?)

• Use url as unique id for articles
• Upsert articles when added
• Add unique slug field for topics, use for app routing
• Take top article per news source from get-news endpoint
• Add logging to get-news

• Deploy to Vercel
• Set up webhook flow to retrieve analysis when article is saved
• Add better logging to get-analysis
• If rate limited, fallback to Open AI

• Limit articles per topic and sort by publish date