Ideate

Ad-hoc research

On /ideate, type anything: a team, a company, a news topic, a cultural moment. Pick the platforms you want posts for, hit Research.

The pipeline: Claude’s native web_search tool (up to 5 hops), scoped to the last 24 hours, returns a JSON blob with:

• A 2–3 sentence summary
• 3–6 trending angles under the topic
• Specific stats with source attribution
• Notable direct quotes
• A list of primary sources (title, url, publisher, date)
• 2–3 draft posts per platform, tuned to each platform’s voice + char limit
• A rough volume estimate (articles/posts seen in the 24h window)

Why it can take 30–60 seconds

The Lambda does multiple web search hops and reads content from each result before it synthesizes. That’s slower than a chat reply but it’s the work we’re paying for. We run the call asynchronously — the SSR route returns a jobId, the client polls every 2 seconds until the job table shows status=done (or times out at 3 minutes).

Pinning

If you click Pin after a result comes back, we create a row in clipunity-ideate-queries. That row is picked up by the clipunity-ideate-cron Lambda every 6 hours, re-runs the research, and writes a snapshot to clipunity-ideate-snapshots.

Each snapshot carries its own volumeEstimate, which powers:

• The sparkline on your /home signal strip
• A trend.alert notification if the latest volume exceeds 2× the prior-run mean
• A briefing.daily notification for the user who pinned the query

Model

We default to claude-haiku-4-5 — fastest, cheapest, and the hook-detection + synthesis are well within its capability. If you want Sonnet for a specific call, flip it under Settings → Preferences → AI model.