Building an AI Friend That Actually Works With You

A friend recommended Nova Seed, so I'm giving it a try to see if it's actually useful. The headline of that Hacker News report was "Create an AI friend and build something together," focusing on working together. I've looked through the project documentation and public shares recently, and I found that the key to this type of product is whether there's a reusable workflow.

Let me explain three terms first. Here, an 'AI friend' is more like giving it a fixed role and boundaries. A prompt is the manual you write for the model, telling it how to speak and act. Context is how much content it remembers in the current chat. Beginners don't need to install local models; just practice with web-based chat tools.

I saw someone share Nova's prompts, which prioritize accuracy, clarity, and emotional safety. This sentence is crucial. Companion-type products, if they only have friendly tones without boundaries, easily lead users into ineffective chats. From app store descriptions, there are many AI chat apps named Nova, mostly emphasizing smart chat, emotional understanding, voice, and emoji interactions. They solve 'talking with me,' not 'helping me get things done.' The key in this niche is running tasks from start to finish.

Here is a beginner-friendly way to set up an AI friend that can work with you:

1. Open your usual chat tool. Web versions of ChatGPT, Qwen, DeepSeek, etc., all work. I've used each for three weeks, a month, and four weeks respectively; the interfaces aren't that different. Click 'New Chat' on the left. Don't ask a question in the input box yet; paste the following instead.

You are 'Little Companion,' an AI friend who works with me. Your goal is to help me turn vague ideas into next-step actions; don't chitchat. Divide every reply into four sections: My Understanding, Next Steps, Risks, Needs Confirmation. If unsure, say you don't know. Keep it in Chinese, use short sentences.

Normal result: It should restate its role and wait for your task. If it starts offering lengthy comfort, the prompt isn't strict enough; add 'No comforting, just next steps.'

2. Input the project card. Don't start with 'Help me write a plan.' Give background, goals, and constraints first. You can input directly.

We are doing a project now: organizing this week's work logs into a weekly report. My goal is for my supervisor to read it, so don't write in a tone meant for myself. Constraints: Under 300 words, must include progress, risks, and next week's plan. I'll send the materials shortly.

Expected result: It waits for materials. If it generates a weekly report already, correct it immediately: materials haven't been provided, so don't generate content yet.

3. Send materials. Beginners often dump chat logs, screenshots, and notes all at once. Suggest organizing into three parts: What did I do this week, Where am I stuck, What to do next week. Even if rough, it's fine. After pasting in the input box, ask:

Please don't write the weekly report yet. First, help me check which of these three types of information is unclear.

This step trains it to do coarse filtering, judging if the materials are sufficient. My previous judgment was that AI is better suited for code screening, test assistance, and data organization, not directly generating production-grade mergeable content. The logic for an AI friend is the same: let it judge if the information is sufficient first.

4. Let it break down tasks. Don't ask 'What's next?' Ask this instead:

Break 'Organizing the weekly report' into 3 steps. Each step under 20 minutes. For each step, tell me what you need to do and what I need to do.

Normal result should be: Step 1 extract progress, Step 2 identify risks, Step 3 compress to 300 words. If it gives empty talk like 'Strengthen communication' or 'Sync more,' ask it to redo it, replacing with actionable steps I can execute immediately.

5. Do only one step per round. Many beginners throw the whole task at once, and the model misses details. The correct approach is this:

Do only the first step now: Extract 5 progress points from the materials below, each under 20 characters.

After seeing the results, manually delete the unimportant ones. Don't let it complete the entire project at once. Humans handle judgment; AI handles organization.

6. Fix the output format. Finally, when letting it write the weekly report, add this:

Output in Markdown, with three headings: Progress, Risks, Next Week's Plan. Do not write a summary paragraph.

This way, you can copy directly into documents. The more fixed the format, the less likely it is to drift.

7. Establish boundaries. Before each chat, you can paste this:

If materials are insufficient, ask me first, don't guess. If involving commitments, budgets, or HR, list risks only, don't decide for me.

This sentence reduces a lot of trouble. The more an AI friend acts like a friend, the more you need to remind it that it's not a decision-maker.

Common pitfalls for beginners, listed here:

Name confusion. Nova Seed on GitHub and Nova AI Friend in app stores might not be the same product. When searching, don't just look at titles; check developers, project docs, and UI screenshots.

Treating it as a toy. Starting with 'Do you like me?' easily turns into emotional chatting. Clearly define the goal first: today we only handle one thing.

Context too long. By the 30th turn, it might forget the constraints from the first turn. Solution is simple: save key rules as a separate block and re-paste them for every new chat.

Letting it fabricate facts. Without materials, it will invent using common sense. Weekly reports, code, finance, medical, legal—none can be trusted directly.

Pasting privacy. ID cards, contracts, client lists, internal system screenshots—don't casually drop them into chat boxes. Beginners most often fail here.

Advantages are clear. Low barrier to entry, no need to understand model principles. Replicable: organize weekly reports today, meeting notes tomorrow. Suitable for long tasks, breaking big things into small steps. It also lowers startup pressure; many people get stuck not knowing where to begin.

Disadvantages are direct. Easily becomes a chat toy. Output quality depends on model capability. Many products share names, leading to varied experiences. If you expect it to automatically complete production-level tasks, you'll be disappointed.

My judgment is that this kind of 'AI friend' is suitable for personal efficiency tools, not critical business systems. You can treat it as an assistant or a sparring partner, but not as a person responsible for outcomes. What beginners really need to learn is breaking vague needs into confirmable next steps.

After mastering this, try changing the role card to 'Code Review Friend' or 'Study Plan Friend,' letting it only handle breakdown and checks, not making final judgments for you.

---

📌 This article is compiled from Hacker News, original source: https://github.com/mas-bandwidth/nova

All rights reserved by the original authors. This is a compilation and independent analysis based on public reporting.