Artifact
A living requirements workspace for cross-functional teams. Paste a messy product idea, get a shared plan with questions for every role.
Most product ideas start messy: a Slack thread, a half-written doc, a sentence in a meeting. By the time they become a plan, half the hard questions have been skipped or answered by whoever spoke last.
Artifact is built for that handoff. You paste a rough request. Andy, the AI collaborator, drafts a structured artifact and routes questions to the right roles. Everyone answers their slice of the same document. An alignment score tracks how close the team is to ship-ready.
The workspace
The main surface is a document, not a chat thread. Content lives in sections. Questions sit inline, colored by role. Switch who you are viewing as and the open questions change with you. Answer one and the alignment score moves.
The demo below is a frozen sample. It behaves like the real product: role filtering, severity labels, and a live score.
Short-form video in the main feedProblem
Feed engagement is drifting to short-form competitors. Users who arrive for social graph updates leave after a few static posts. We need a native video surface that keeps people in-product without breaking the existing feed contract.
What success metric defines “video is working” in the first 90 days - watch time, retention, or creator upload rate?
Technical approach
Video will ship behind a feature flag. Phase 1 reuses the existing media pipeline for playback and adds an encoding queue for creator uploads. CDN cost estimates assume 40% of daily actives watch at least one clip.
Do we need a new moderation SLA for video, or can text-policy tooling cover the first release?
Ship with human review on report + existing text classifiers on captions. Full video ML moderation is a follow-up quarter.
Experience
Creation should feel closer to a camera app than a form. Capture, trim, caption, post - under 60 seconds for a first draft. Playback is full-bleed in the feed with sound off by default.
Rollout
Internal dogfood for two weeks, then 5% of creators, then general availability. Kill switch lives in the feed config service.
Which creator cohort is the dogfood group, and who owns the go/no-go checklist?
Top 2% creators by weekly posts in three markets. PM owns go/no-go with Eng + Trust sign-off.
Why it exists
Cross-functional work rarely fails because the idea is bad. It fails in the gaps between people. Product writes a one-pager. Engineering finds the load path missing. Design discovers the happy path assumes three steps users will not take. Legal arrives after the mockups. Support learns about the feature from a confused ticket.
The meeting that was supposed to refine the plan becomes the meeting that invents it. I wanted a surface where every stakeholder could see their part of the same artifact, and where gaps were questions with owners, not vibes in a shared doc.
Assemble the team
Generation takes a few seconds. Instead of a spinner, Artifact asks you to assemble the team while the structure is drafting. Dead time becomes a decision: who needs a seat at this table?
Structure comes first, content only. Questions come next, tailored to the roles you picked. That split keeps the first draft useful even before anyone has answered anything.
Assemble your team
While Andy drafts the structure, you choose which perspectives get questions.
3 roles will receive questions
Roles as lenses
The unit of work is not a blank PRD and not a free-form chat. It is a role-routed question embedded in the document. Product sees strategy gaps. Engineering sees feasibility gaps. Design sees experience gaps. Same artifact, different lens.
Andy is the facilitator. He drafts, asks, and rewrites when answers come in. The point of giving him a face is not decoration. It is to make the system feel like a colleague coordinating the room, not a form that emits text.
Alignment
“Are we aligned?” is usually a feeling. Artifact turns it into a score: drafted sections, answered questions, and open blockers from an audit. Status moves from Draft to In Review to Ready as the number climbs.
Scrub the control below to see how the thresholds work. In the live product the score updates as people answer.
- Draft · under 30
- In Review · 30 - 79
- Ready · 80+
Score is not a vibe check. It weights drafted sections, answered questions, and open blockers from the alignment audit.
How it's built
Next.js for the app, LLMs for structure and questions, Prisma on Postgres for persistence, share tokens for async collaborators. The live product takes your own Anthropic, OpenAI, or SpaceXAI key. The key stays in the browser and is only used to call the provider.
- Next.js
- LLMs
- Prisma
- Postgres
Artifact is a prototype of a simple bet: messy input in, multiplayer structure out, progress you can point at. If you want the full loop, open the live product and bring a key.