DEV Community: Dror Wayne

The guide to building apps with AI

Dror Wayne — Wed, 12 Feb 2025 12:24:39 +0000

Anyone can build apps using AI - here's what you need to know.

Click here for the full guide by Dan Leshem

Prompt Engineering? You're doing it wrong.

Dror Wayne — Tue, 04 Feb 2025 08:46:47 +0000

*I don’t believe you need to become a “prompt engineer” to succeed with AI coding - by contrast, all you need to do is write very clear, simple specs. *

Some of the most common pitfalls that my team has been noticing, come down to users overcomplicating their prompts and confusing the AI.

Roundabout wording, with long sentences, can make it unclear which files should be edited and why they’re being mentioned. Instead of long paragraphs, I recommend structuring clear sentences, using a formula such as “Given, Where, When”.

Some people fall into **the trap of combining catch-alls with specifics **in their prompts. Each prompt should be either specific or general. It’s also OK to start off with a general query - let’s say, search the repo for dependencies on a component, before you write the specs to change said component. Or if you prefer, after giving a task, follow up with a catch-all question such as, “Now that we have changed ____, what else can or should be edited in this file to match?”

The most ridiculous “prompt engineering” that I see is when users end up contradicting themselves. “Fix this bug but don’t change any of my code”. Well, how is the bug going to be fixed?

**Also - you don’t need to start every prompt with “You are a senior software developer” anymore. **It’s 2025. (At least not when you’re using Fine… I can’t speak for every platform out there).

Remember, in most platforms you can add overriding rules or instructions which form a part of the agent’s flow with every task. In Fine it’s project based - you can add custom instructions (style, language, commenting etc) which apply to all tasks on the repo, but don’t confuse the agent when reading the specific task.

Why you're not succeeding with AI coding

Dror Wayne — Mon, 03 Feb 2025 07:04:00 +0000

Interesting piece of research that reinforces what I’ve been saying for a while about AI Coding

Annie Vella studied the correlation between the frequency that developers use AI coding tools and the impact it has on their productivity, with clear results: you need to use the tools at least 2-3 times per week to start seeing significant productivity gains.

This reflects my own anecdotal experience and the trends I’ve noticed talking to developers whilst building Fine. To succeed with AI coding, there are certain skills you need to build, adjustments to make, approaches to learn.

A developer can save hours with the help of AI, but do so they’ll need to learn how to triage the tasks they delegate to AI, how to write specs, how to review the AI’s code efficiently and overall how to best collaborate with AI agents.

Unfortunately, if you tried coding with chatGPT in 2023 and haven’t tried using AI again since because it didn’t deliver - it’s unlikely that you’ll be able to succeed.

My recommendation for developers getting started with AI is to take it easy, but be consistent. Try using an AI Agent to resolve one ticket a day for 2 weeks. Some will fail, some will get 80% of the way, some might be very impressive one-shots. But by the end of two weeks of daily use, you’ll have the skills to make the most of AI in your daily work - for your own benefit.

If you're down - head to fine.dev where you can try out AI coding with different LLMs and connect to your codebase to see where you can save time.

Replit vs Cursor

Dror Wayne — Sun, 26 Jan 2025 15:03:02 +0000

Why choose Cursor over Replit?
Why choose Replit over Cursor?
How you can connect cursor and replit!
Why Fine beats both Cursor and Replit
Here's the article.

Shipping UX improvements

Dror Wayne — Wed, 22 Jan 2025 12:30:00 +0000

*We've been shipping a lot of UX improvements recently. Here's how we do it.
*
1) Maintain a backlog of tasks. Don't let ideas and suggestions get lost in slack channels - list them all. We use Linear.
2) Triage tasks to decide which are most suited to AI. I can go into this more, but think tasks that would take up to 3 hours, touch up to 3 files and have clear acceptance criteria.
3) Add a label to those tasks ("AI-ready")
4) When you have the time, come back and pick 2-3 tasks to give to the AI
5) Run all three
6) Review the logs, visually verify the changes with Live Previews and make any changes needed
7) Create PR

𝐓𝐡𝐞 𝐭𝐚𝐬𝐤𝐬 𝐭𝐡𝐚𝐭 𝐮𝐬𝐞𝐝 𝐭𝐨 𝐚𝐝𝐝 𝐮𝐩 𝐭𝐨 80% 𝐨𝐟 𝐨𝐮𝐫 𝐭𝐢𝐦𝐞 𝐚𝐭 𝐦𝐲 𝐩𝐫𝐞𝐯𝐢𝐨𝐮𝐬 𝐬𝐭𝐚𝐫𝐭𝐮𝐩, 𝐧𝐨𝐰 𝐭𝐚𝐤𝐞 20% - 𝐛𝐞𝐜𝐚𝐮𝐬𝐞 𝐨𝐟 𝐭𝐡𝐞 𝐰𝐚𝐲 𝐰𝐞'𝐫𝐞 𝐮𝐬𝐢𝐧𝐠 𝐀𝐈.

Dror Wayne — Tue, 21 Jan 2025 12:29:00 +0000

𝐓𝐡𝐞 𝐭𝐚𝐬𝐤𝐬 𝐭𝐡𝐚𝐭 𝐮𝐬𝐞𝐝 𝐭𝐨 𝐚𝐝𝐝 𝐮𝐩 𝐭𝐨 80% 𝐨𝐟 𝐨𝐮𝐫 𝐭𝐢𝐦𝐞 𝐚𝐭 𝐦𝐲 𝐩𝐫𝐞𝐯𝐢𝐨𝐮𝐬 𝐬𝐭𝐚𝐫𝐭𝐮𝐩, 𝐧𝐨𝐰 𝐭𝐚𝐤𝐞 20% 𝐚𝐭 𝐅𝐢𝐧𝐞 - 𝐛𝐞𝐜𝐚𝐮𝐬𝐞 𝐨𝐟 𝐭𝐡𝐞 𝐰𝐚𝐲 𝐰𝐞'𝐫𝐞 𝐮𝐬𝐢𝐧𝐠 𝐀𝐈.

We've offloaded all our high-cognitive load, limited system-knowledge tasks to AI Agents (for more on that see my previous substack).

When you add it all up, it's taking 1/4 of the time it used to take - because of how much work the AI Agents do for us.

Once the specs are written, the developer involvement is down to 4 steps - of which 3 are just one click each.
Delegate

the issue (in the web app, Linear or Slack).
Approve the implementation plan
Review the code changes, logs and preview (this is where it's more than one click) a. Approve b. Make quick manual revisions c. Prompt Agent to revise code
Create PR

The Agent is doing everything else 𝒐𝒏 𝒊𝒕𝒔 𝒐𝒘𝒏.

So all of those UX/UI improvements, minor bugs to squash and simple features to develop - are done by the first coffee break of the day, leaving us to spend time on planning the future of the product and the new challenges we are undertaking.

P.s. A key factor is identifying which tasks are best offloaded to AI. I shared practical advice, key criteria and an AI script with Fine users and my substack subscribers. If you’re interested in reading it, comment below and I’ll send you the link.

AI Coding isn't just replacing developers

Dror Wayne — Mon, 20 Jan 2025 16:05:18 +0000

When we set out to create Fine, our plan wasn’t to “replace developers” or to get AI to write code for the sake of AI writing code.

We wanted to create an "AI teammate" experience of collaborating with AI like you'd collaborate with a colleague developer.

Well, developers need their own dev environment - with the source code, build tools, dependencies and customizations.

If we could create this AI teammate, with a dev environment of its own, in the cloud, the benefits would be awesome:
Working i

n parallel on multiple tasks
Switching between tasks with a much lower overhead
Get tasks done (or started) on-the-go
AI can work for you asynchronously, not just whilst you’re involved
Easy to test and verify the AI generated code

We’ve been working on Fine for a long time but we’re now at a stage where we can look back and see that we achieved this goal and developers are able to use their AI Teammate in their day-to-day work and not just for hobbies.

We’ve got a full roadmap of features, integrations and improvements coming over the next few months, to make your AI Teammate as versatile and useful as possible. If you’re interested in trying out coding with AI for your startup - check out fine.dev

𝐅𝐢𝐧𝐞 𝐯𝐬 𝐂𝐥𝐚𝐮𝐝𝐞 & 𝐂𝐡𝐚𝐭𝐆𝐏𝐓 - 𝐖𝐡𝐚𝐭'𝐬 𝐭𝐡𝐞 𝐝𝐢𝐟𝐟𝐞𝐫𝐞𝐧𝐜𝐞 ❔

Dror Wayne — Mon, 20 Jan 2025 07:00:00 +0000

A question I've been asked regularly, so here's a breakdown on the practical differences for you as a developer.

🗨️ With Claude and ChatGPT, there's no context. If you give them a task, you;ll get a generic response, because they don't know your code.
🗨️ Alternatively, you copy and paste your code into the chat. Apart from being complex and time-consuming, you can only really do so with one file at a time.
🟣 Fine is connected to your GitHub and indexes your codebase, so it can quickly find the relevant context and write accurate, applicable code. It can edit multiple files at the same time, ensuring they work together.

🗨️ If you copy code into Claude or ChatGPT, the AI may introduce new bugs, by editing or deleting existing code - and there's a good chance you don't spot the changes.
🟣 Fine shows clear diffs and a Line Change Summary. It also creates an implementation plan, telling you what it's going to do, does exactly that, and tells you what it's done.

🗨️ To test code generated by Claude or ChatGPT, you're going to need to copy it back into your IDE to run.
🟣 Fine allows you to run, test and preview the code in the browser, whether you're accessing from your mobile or laptop.

These are just 3 of the major differences between working with the LLMs in their basic format designed for chatting and working with Fine's Agents, powered by the same LLMs, but built for developing.

For AI Coding - use an AI Coding Agent. Not an AI Chat.

𝐖𝐡𝐚𝐭 𝐤𝐢𝐧𝐝 𝐨𝐟 𝐢𝐬𝐬𝐮𝐞𝐬 𝐬𝐡𝐨𝐮𝐥𝐝 𝐲𝐨𝐮 𝐝𝐞𝐥𝐞𝐠𝐚𝐭𝐞 𝐭𝐨 𝐲𝐨𝐮𝐫 𝐀𝐈 𝐜𝐨𝐝𝐢𝐧𝐠 𝐚𝐠𝐞𝐧𝐭?

Dror Wayne — Sun, 19 Jan 2025 12:25:54 +0000

One key to successfully adopting AI into your workflow is knowing to which tasks AI is suited.
❌ Tasks which are too complex/big will fail
❌ Tasks which are too small don't bring you benefit

Working with a small team and heavily adopting AI to help us ship faster and build our startup, I've learned to triage the tasks I delegate to AI as follows:
✅ Touches max 3 files
✅ Clear success criteria
✅ Single repository
✅ No gotchas
✅ Would take up to 3 hours without AI

When I go through our development backlog looking for these sorts of issues, what I find we're left with tends to include
✅ All minor bugs
✅ UI tweaks
✅ Chore tasks (e.g. adding proper error handling)

In summary, what we call "low-surface context" tasks, where the path to resolution is clear and straightforward.

P.s.
🤖 I've built a script that goes through your Linear and triages your tasks according to these criteria, to add an AI-ready label. I shared it with my substack subscribers. If you're interested, comment below and I'll send you the link.

New Fine Updates - Project Instructions & More

Dror Wayne — Fri, 03 Jan 2025 07:00:00 +0000

By popular request, we’ve added the ability to define your own set of custom instructions or rules for the AI Agent. Head to Projects > Project Settings > Instructions.

You can use Project Instructions to set coding style preferences and conventions, include external documentation and style guides.

Instructions will be used by the Agent for every task on the Project, from all users in the workspace.

You can apply different instructions to different projects and if you have shared instructions that you would like to apply to multiple projects, make sure to copy them into each one

MIND-BLOWING AI Agents Fix Code in SaaS platform - FASTEST way to improve software - UNEDITED

Dror Wayne — Wed, 01 Jan 2025 13:59:19 +0000

AI Coding Agents that run and test the code

Dror Wayne — Wed, 25 Dec 2024 11:25:20 +0000

Fine just released two features on Christmas and it's set to be an awesome 2025

Fine, the AI Coding Agent for Startups that connects to your GitHub and completes dev tasks for you, just dropped AI Sandboxing and Live Previews. Here's what it means:

Fine's AI Agents create a new branch to edit/create files in your codebase, using a unique Cloud Dev Environment for each task. Now, you can run the code in that same dev environment - no need to load up your IDE.

Watch your code compile in the console and if there are any errors, just copy and paste the logs into the chat for the Agent to fix.

You can also open a Live Preview to visually verify your changes. Share the Preview URL with colleagues to get feedback and for backend work, test with tools such as Postman.

Why this is important for developers

Imagine a world in which AI not only takes your specs and writes the code, but can run and test the code, compare it to the specs and fix itself. Where AI not only writes code that looks right and makes sense to the LLM’s logic, but verifies that it actually works - and meets the exact requirements you defined.

AI Sandboxing is not just another Cloud Dev Environment - it's the next step in achieving this.

With the help of your feedback on the current version of Live Previews, we’ll soon be setting our sights on releasing a complete self-assessing, self-improving AI coding agent - the first ever.