Cold Start

Cold start is the first real interaction a user has with an AI product. It is the moment the mental model of a non deterministic system starts to take shape around the users own data instead of abstract theory. Good cold starts drop the user straight into their actual codebase their actual documents or their actual next task. This forces the four onboarding jobs to resolve in context. Capability bounds become obvious when the AI interacts with real constraints. The interaction model is taught through familiar shortcuts applied to familiar files. Success looks like a useful edit on work the user already cares about. The first prompt happens naturally because the context suggests it. Cursor executed this perfectly in its 2023 launch by making the cold start an immediate request to open a real folder. The product stays out of the way until the users own work loads then surfaces AI features exactly where they become useful. The mental model sticks because it was built on reality from the first second.

Cold start is not a demo environment populated with clean sample data. It is not a multi step tutorial that ends in an empty prompt bar. It is not a carousel of feature explanations or a form that demands profile information before any value appears. Those patterns all create distance between the user and their real work. They let the user practice on fake problems and then abandon the product when real problems feel different. A cold start rejects safety theater in favor of honest collision with the users actual world even when that world is messy. Early design AI tools in 2023 failed exactly here by shipping beautiful sandboxes full of placeholder wireframes and stock photos. Users had fun tweaking perfect demos then watched the same model choke on their real client files full of custom components brand constraints and messy layers.

Concrete examples show how powerful the pattern becomes when executed well. Cursor remains the gold standard for code tools. Instead of a guided tour the product opens a native dialog that insists on a real project. Once loaded the AI suggests improvements to code the user shipped last month. The first acceptance loop happens on production code not a tutorial. This teaches the limits fast. The model might decline to touch certain legacy files or it might generate tests that reveal gaps in the users understanding. Granola applied cold start to meetings. The onboarding flow is literally one line connect your calendar. The product then waits for the users next real call and joins it. The first output the user sees is notes from a meeting they actually attended. No practice runs. No fake transcripts. The success state is immediate because the context is personal. Perplexity uses prompt suggestions as a lightweight cold start. The homepage loads with timely questions that pull live information. Clicking one produces a complete answer page with sources and related queries visible. The user learns the entire interaction model from one real query.

Claude.ai loads a page of rich example prompts on first visit. Each example is detailed enough to show tone length and structure. One click opens a live chat with that prompt pre filled. The user can edit and send or just watch the output generate. This collapses the time to first success dramatically. Linear AI avoids a dedicated cold start screen entirely by embedding AI inside flows the user already runs every day. When creating an issue the AI offers to draft the description based on similar past work in the project. The feature feels like an extension of the product not an add on. ChatGPT in late 2022 used its famous three column layout as cold start. Examples on the left showed what you could ask. Capabilities in the middle set the scope. Limitations on the right set honest expectations. That page taught more in ten seconds than most current products teach in five minutes.

We have also seen strong cold starts in creative tools. Midjourney began inside Discord channels where new users saw community prompts generating images in real time in 2022. They could immediately remix those prompts and see variations. The mental model formed through live observation rather than instruction. In design tools the pattern appears in products like Galileo AI that let users upload their own brand assets on first run. The AI then generates new screens that respect those exact colors and typography instead of generic material design. This beats any tool that starts users in a blank state with preset styles that have no relation to real client work. Adobe Firefly integrated into Photoshop follows the same logic. Select an object in your own document and generative fill operates on your actual layers colors and composition instead of a sterile test canvas.

Use a cold start when your product can ingest the users existing data without high friction. Code assistants writing copilots data tools and design enhancers all benefit. The pattern works best when the users data makes the outputs obviously personal and when the tool ties to recurring behaviors like meetings commits or content reviews. The event triggered version Granola ships removes the need for the user to even remember to open the product. Avoid cold start when the setup cost is too high or when the user likely has no data yet. A brand new workspace or empty repo creates a bad cold start. In those cases use prompt examples or hybrid approaches that offer a temporary demo then transition quickly to real files. Never use it for completely general tools that have no persistent user context like basic chat interfaces. Those win with strong default prompts and visible capability bounds instead. Forcing a cold start on a pure research oracle or open ended image generator just adds pointless steps.

The best cold starts make the user forget they are being onboarded at all.